You are on page 1of 40

Received: 22 June 2018 Revised: 12 February 2019 Accepted: 24 February 2019

DOI: 10.1002/oca.2496

L I T E R AT U R E R E V I E W

Discrete-time systems sliding mode switching hyperplane


design: A survey

José Darío Luis-Delgado1 Basil Mohammed Al-Hadithi1,2 Agustín Jiménez1

1
Intelligent Control Group, Centre for
Automation and Robotics UPM-CSIC, Summary
Madrid, Spain As regards to variable structure control with sliding mode (VSC-SM), it is known
2
Department of Electrical, Electronics,
that designing sliding surface is a critical task because sliding surfaces must
Automation and Applied Physics, School
of Industrial Design and Engineering, guarantee that the steady state of the controlled system achieves some desired
Technical University of Madrid, Madrid, performance including global (or global asymptotic) stability. This problem has
Spain
been extensively analyzed for a wide class of systems: scalar, multivariable, lin-
Correspondence ear, nonlinear, time-variant, discrete-time system, etc. The main purpose of this
José Darío Luis-Delgado, Intelligent paper is to present a thorough survey about the main approaches related to the
Control Group, Centre for Automation and
Robotics UPM-CSIC, 28500 Madrid, Spain.
design of discrete-time switching surfaces applied to VSC-SM. Research works
Email: jd.luis@alumnos.upm.es covering the design of sliding surfaces for the aforementioned systems are con-
sidered and some illustrative examples are also included to explain some of the
Funding information
Spanish Ministry of Science, Innovation design methodologies.
and Universities, Grant/Award Number:
DPI2017-86915-C3-3-R K E Y WO R D S
discrete sliding mode, nonlinear control, switching surface, variable structure control

1 I N T RO DU CT ION

As far as variable structure control with sliding mode (SM) is concerned, it is well known that the design procedure is
implemented in two steps or phases; the first one, the design of the switching function or the sliding surface; and the
second one is specifying the control law. Furthermore, the first step deals with the closed-loop sliding phase and the
second one with the reaching phase. Due to VSC-SM corresponding to the steady state of the closed-loop system, then
specification of switching functions parameters is critical because the performance of the controlled system depends on
it. If a sliding surface is not well designed, then the overall performance of the controlled systems could be poor and
undesirable behavior may arise.
The main differences between the design of sliding surfaces in discrete-time domain compared with the
continuous-time one are, basically, two. The first one is that the dynamic of the ideal SM must be designed to fulfill either
classical discrete-time or continuous-time stability criteria, such that, for linear systems, the eigenvalues of the equiva-
lent controlled dynamics must be located inside the unitary circle for discrete-time systems instead of the left side of the
poles plane for the continuous ones, and for nonlinear system design, the right corresponding Lyapunov stability theorem
must be applied. The second difference is that ideal SM dynamics in discrete time cannot be achieved, ie, oscillations

Abbreviations: ANA, antinuclear antibodies; APC, antigen-presenting cells; IRF, interferon regulatory factor.

Optim Control Appl Meth. 2019;1–40. wileyonlinelibrary.com/journal/oca © 2019 John Wiley & Sons, Ltd. 1
2 LUIS-DELGADO ET AL.

around the sliding surfaces always exist.1-4 Hence, the reduction of this oscillations amplitudes is a main design objective
for discrete-time VSC.
In general, the calculation of the sliding surfaces for variable structure controllers in discrete-time systems has been
made in a similar way to the case of continuous ones, where it is considered that the controlled system will exhibit an
ideal SM, which is never achieved in discrete time. However, this direct translation of the techniques from continuous to
discrete time does not consider many relevant aspects such as the existence of a sliding band around the sliding surface
within which the state of the controlled system oscillates.1-4 In addition, both the dynamics of the nonideal SM and the
width of the sliding band depend on both the dynamics of the system, the sampling frequency,2,5,6 and the quantification
error.7,8 Therefore, the design of discrete-time VSC (DVSC) controllers is a problem that remains unsolved at all or is
solved partially, which is shown by the large number of works described in this survey. If we consider the number of
recent research works related with discrete-time DVSC that are continuously and frequently published,9-11 then it can
be said that this problem continues to be relevant and of great interest to be solved. For this reason, the authors have
decided to prepare this survey, which corresponds to a comprehensive compilation of the most relevant works published
so far, which has involved a great effort in conducting an exhaustive search and a detailed analysis of the works found to
elaborate.
The main aim of this work is to present a brief description of the main research areas related to the design methods of
discrete switching surfaces. For this purpose, a compilation of the main historical issued works has been revised, and a
significant set of approaches has been chosen and presented in this work. The criterion for the selections of the develop-
ments is based on (i) how much relevant or novel is the work under study; (ii) if the proposed design method is efficient
and easy to implement; and (iii) if the research work establishes a new research area such that LMI, min-max control, etc.
Finally, some numerical examples are included to illustrate some of the explained methods. Although the main proposal
of this work is to analyze and discuss design methods applied to discrete-time systems, some continuous time switching
surface design approaches are also incorporated.
This paper is organized as follows. In Section 2, the general sliding hyperplane designing problem is stated. In Sections 3
and 4, some relevant approaches related to the linear and nonlinear systems sliding surface design process respectively
are presented. Section 5 covers the chattering elimination problem. Section 6 is dedicated to illustrative examples. Finally,
a remark summary and the conclusions are detailed in Sections 7 and 8, respectively.

2 PROBLEM STAT EMENT

The objective of designing sliding surface is, basically, to guarantee that the closed-loop system is globally stable or glob-
ally asymptotically stable. This problem has been covered since the earliest works about VSC-SM.12-17 Most of them deal
with continuous linear systems. In the middle of the 80s, some detailed studies about DVSC appeared and, consequently,
an increasing number of research works engaged in analyzing and specifying discrete-time sliding surface1,2 have been
carried out. Regarding linear system sliding surface design, there are two mainly research lines based on classical linear
system designing methods13 : the first one applies method of arbitrary eigenvalues assignment14,15,18-21 and the second one
includes optimal control procedures, such linear quadratic regulator (LQR) or, more recently, linear matrix inequalities
(LMIs).22-27 Other proposed techniques include, for example, linear Lyapunov equation approach,28 H∞ robust scheme,29
or output feedback min-max controllers methodologies.30-32 Some developments deal with particular types of VSC such as
integral VSC-SM21,33-36 and other approaches have proposed the design of nonlinear surfaces to improve transient perfor-
mance of discrete-time uncertain systems37 or to obtain high speed response in delayed discrete-time linear systems.38 A
recent work39 introduces a generic and flexible design proposal applicable to discrete-time mulitvariable linear systems.
On the contrary, for nonlinear systems, a reduced number of proposals have been carried out. Nevertheless, the problem
has been analyzed from the middle of the 80s (see, for example, related studies3,40-45 and more recent works46-48 ).
As in every controlled system, VSC is applied to stabilize and drive the outputs of a process or plant x(k) or y(k) to some
desired values adjusting the control signals u(k). A particular feature of the VSC (see Figure 1) is that these controllers are
built with two main parts: the first one is a multivariable function known as switching surface (highlighted by the green
box in the picture) and a bank of control law blocks or a big multivariable function that made use of the output of the
switching surface to obtain the desired control signals u(k) to the plant. As it was previously mentioned, the aim of this
survey work is to give a brief review of the main research works and approaches that have been proposed to define the
switching surface. The VSC control philosophy is established in the next problem statement.
LUIS-DELGADO ET AL. 3

FIGURE 1 General scheme of a closed-loop system controlled via discrete-time variable structure control (DVSC) [Colour figure can be
viewed at wileyonlinelibrary.com]

Problem 1 (Sliding surface design problem).


Given (x, u, k), a controllable and observable discrete-time system is assumed to be known and completely
defined by

x(k + 1) = 𝑓 (x(k), u(k)) , (1)

where x ∈ X ⊂ ℝn is the state vector, u ∈ U ⊂ ℝm is the input vector, and 𝑓 ∶ ℝn×m → ℝn is a smooth function.
If there exists a control law u(k) such that, in some finite time kr , it is guaranteed that s(x(k + 1)) = s(x(k)) = 0 for
some function s ∶ X → ℝm , then find s(k) = 0 such that, ∀k > kr , (x, u, k) is asymptotically stable.
As it is stated in Problem 1, the control law u(k) must be assumed to be specified such that any state is steered from the
initial state x(0) toward the sliding surface and, once the state has reached s(x(k)) = 0, then it remains indefinitely on it,
ie, the control law must be specified such that an SM is achieved. Hence, given an equilibrium state xe ∈ , where  is
the state subspace where sliding or quasi-SM4 is achieved and defined as follows:

 = {x ⊂ X ∶ s(x) = 0} . (2)

Hence, the design problem consists in finding the switching function s(x(k)) such as the closed-loop system, xe should be
stable in the Lyapunov sense. Hence,

lim x(k) = xe , x(kr ) ∈ . (3)


k→∞

Thus, it is not only necessary that the state dynamic to be stable on the sliding surface  ⊂ X but also the state is needed
to be driven toward some equilibrium state xe ∈ . In general, the sliding surface is defined as the intersection of m planes
si (x) such that rank[si (x)] is n − m, ie,

m
s(x) = Ci x(k), CiT ∈ ℝn , (4)
i=1

so the SM only occurs in the intersection of the Ci x(k) planes. Different schemes of how to deal in the reaching phase with
this multiplanes surface can be found in other works.16,17,20,49 On the other hand, if (4) is written as the matrix product

s(x) = Cx(k), (5)

then it is clear that  = N(C), ie, the SM corresponds to the null space of C, where C ∈ ℝm×n is defined as follows:
[ ]
CT = C1T C2T · · · Cm
T
. (6)
4 LUIS-DELGADO ET AL.

In the work of Slotine and Li,49 the authors described the sliding surface as follows.
Proposition 1. Given ̃ x = x − xd as the state vector tracking error of the system (1), then the sliding surface s(t) ∶ ℝn−1 →
ℝ is defined as follows:
s(x, t) = C ̃
xT = 0, (7)
where C ∈ ℝn−1×m ≠ 0.
The condition s(x, t) = 0 only holds iff ̃x = 0, ie, s(x, t) = 0 iff x(t) = xd (t). Hence, finding a control law that solves
̃
x = 0 is equivalent to solve the mth linear differential equation given in (7). Furthermore, in the work of Slotine and Li,49
the following sliding surface is proposed:
( )n−1
d
s(x, t) = +𝜆 ̃
x = 0. (8)
dt

There exists a wide number of research works of process control where the switching surface used is in the form given
in (8) such as in thermal mixing tank processes,50 drum water level in an industrial boiler unit,51 control of switched
reluctance motor,52 and car suspension systems.53,54
Finally, it is known that analyzing SM dynamics is not an easy task because the control laws for variable structure
systems are not continuous functions, ie, they are not infinitely differentiable. Consequently, approximating methods for
the sliding dynamics are needed, such as the equivalent control approach,12 which describes the dynamics of the ideal
SM assuming that the control law is a continuous function that guarantee the sliding mode.
The next two sections present the most relevant sliding surface design methodologies, which the authors have found
in the literature. In order to give a more detailed and simple description, the working methods are presented in two large
groups: the techniques related to LTI systems and those dealing with nonlinear systems.

3 DI S C RETE-T IM E LT I SY ST E MS SLIDING SURFACE DESIGN

The general aim of designing sliding surfaces applied to MIMO LTI systems can be defined as follows.
Definition 1. Given a discrete-time linear system with state space dynamics
x(k + 1) = Φx(k) + Γu(k), (9)
where u ∈ ℝ and x ∈ ℝ are the input signal and state vector, respectively, and one sliding manifold  defined in (5),
then it is possible to find an equivalent control law uequ (k) such that the following condition can be fulfilled:
s (x(k)) = s (x(k + 1)) = 0. (10)
That is,
uequ (k) = −(C Γ)−1 CΦx (k) , (11)
then the SM dynamic is given by
xe (k + 1) = Φe x(k)
[ ] (12)
Φe = In − Γ(C Γ)−1 C Φ.

From (12), it can be seen that the ideal SM dynamic or equivalent SM dynamic xequ (k) is linear, therefore x(k) = 0
is a stationary state in the sense specified in (3). Indeed, if (3) holds, then the stationary state x(k) = 0 is Lyapunov
stable. Consider the characteristic polynomial of (12) as follows:
XΦe (𝜆) = (z − 𝜆1 ) (z − 𝜆2 ) · · · (z − 𝜆n ) , (13)
where 𝜆i , i = 1, 2, … , n, are the eigenvalues of Φe denoted as Λ(Φe ), ie, 𝜆i , ∈ Λ(Φe ). Hence, the problem of designing
the sliding surface  is reduced to specify matrix C so that XΦ (𝜆) being strictly Schur,55 ie, ||𝜆i || < 1 for every 𝜆i in Λ(Φe ).
There exists an extensive published development about designing sliding surfaces applied to discrete-time LTI systems,
where the most relevant ones are as follows:
• arbitrary eigenvalue assignment;
• integral SM surfaces;
• surface specification for LTI systems via Lyapunov stability analysis approach;
LUIS-DELGADO ET AL. 5

• optimal control approaches;


• flexible sliding surface design considering nonideal SM dynamic approach.
In the next sections, these schemes are briefly explained.

3.1 Arbitrary eigenvalue assignment


One of the first approaches applied to solve the problem of designing sliding surface is the eigenvalue assignment
methodology.12-14 These methods consist in specifying a linear sliding surface such that the eigenvalues of the equivalent
system matrix can achieve an asymptotically stable SM dynamics. The following sections explain some of the most impor-
tant research works. Some research works of DVSC are applied to industrial processes where a pole placement approach
is used for the design of a discrete-time switching surface, such as for chemical processes,56 for liquid propellant rocket
motor systems,38 and for a linearized model of the aircraft.21
The following sections summarize the main arbitrary eigenvalue assignment approaches that have been developed so
far, which are the canonical system dynamics representations approach, the Ackermann equation–based methodology,
and the flexible sliding surface design considering nonideal SM dynamic.

3.1.1 Canonical system dynamics representation approach


It is known that some form of representation of the state space systems dynamics are more helpful to the task of analysis
and controllers design. For the VSC specifying procedures, some canonical representations have become into stan-
dards representation because they carry out more simple design methods, for example, reduced, controller, and normal
canonical forms are widely implemented as it is explained in the next sections.

Controller canonical representation


It is known that the design of sliding surfaces for single-input–single-output (SISO) linear systems becomes easier if the
state space representation of the system is given in the Luenberger canonical controller form.2,16,17,42
Definition 2. Given (x, u), a discrete-time SISO linear system, whose dynamics are described in (9). The system is
said to be represented in Frobenius or Luenberger canonical control form57 or simply controller canonical form if the
matrices Φ and Γ have the following structure:
⎡ 0 1 0 ··· 0 ⎤ ⎡0⎤
⎢ 0 0 1 ··· 0 ⎥ ⎢0⎥
Φ=⎢ ⋮ ⋮ ⋮ ⋱ ⋮ ⎥,Γ = ⎢0⎥. (14)
⎢ 0 0 0 0 1 ⎥ ⎢⋮⎥
⎢a a ⎥ ⎢1⎥
⎣ n n−1 an−2 ··· a1 ⎦ ⎣ ⎦

If the dynamics of system (9) is not represented as the controller canonical form (14) and if the pair (Φ, Γ) is controllable,
ie, rank[Wc (Φ, Γ)] = n, where Wc (Φ, Γ) ∈ ℝn×n is the controllable matrix computed as follows:
[ ]
Wc (Φ, Γ) = Γ ΦΓ Φ2 Γ · · · Φn−1 Γ , (15)
then a similarity transformation x̂ (k) = Tx(k) can be found such that
̂ x(k) + Γu(k),
x̂ (k + 1) = Φ̂ ̂
̂ = T −1 ΦT and Γ̂ = T −1 Γ in the canonical form described in (14) and where T ∈ ℝn×n is obtained as
with Φ
( )
T = Wc (Φ, Γ) Wc−1 Φ,̂ Γ̂ . (16)

Let us define a linear sliding surface as in (5) such that the sliding manifold  = N(C) is defined as follows:
 = Cx(k) = 0, (17)
[ ]
where C = c1 c2 · · · cn . Without loss of generality, let cn = 1. Therefore,

n−1
xn (k + 1) = − ci xi (k). (18)
i=1
6 LUIS-DELGADO ET AL.

From the equivalent control methodology, the ideal SM dynamics is given in (12), therefore the ideal reduced-order SM
dynamics is as follows:
⎡ x1 (k + 1) ⎤ ⎡ 0 1 · · · · · · 0 ⎤ ⎡ x1 (k) ⎤
⎢ x2 (k + 1) ⎥ ⎢ 0 0 · · · · · · 0 ⎥ ⎢ x2 (k) ⎥
⎢ ⋮ ⎥ = ⎢ ⋮ ⋮ ⋱ ⋱ ⋮ ⎥⎢ ⋮ ⎥, (19)
⎢ x (k + 1) ⎥ ⎢ 0 0 · · · 0 1 ⎥ ⎢ xn−2 (k) ⎥
⎢ x (k + 1) ⎥ ⎢ c c · · · c
n−2
⎥⎢ ⎥
⎣ n−1 ⎦ ⎣ 1 2 n−2 cn−1 ⎦ ⎣ xn−1 (k) ⎦

where the characteristic polynomial is estimated as follows:


XΦ′ = zn−1 + c1 zn−2 + c2 zn−3 + · · · + cn−1 . (20)
Providing that Λ(Φequ ) must be set to some desired values 𝜆d ∈ ℝn−1 , then the resulting characteristic polynomial XΦequ (𝜆)
is defined as follows:
XΦ′ (𝜆) = (𝜆 − 𝜆1 ) (𝜆 − 𝜆1 ) · · · (𝜆 − 𝜆n−1 )
= 𝜆n−1 + 𝛾1 𝜆n−2 + 𝛾2 𝜆n−3 + · · · + 𝛾n−1 . (21)

Hence, the coefficients ci of (20) can be directly found from (21).


For multi-input linear systems where u ∈ ℝm , the similarity transformation can split the system dynamics into
m subsystems,16,42,57,58 where every subsystem is represented in the controller canonical form. Thus, a multiple-input–
multiple-output (MIMO) controller canonical representation can be made where the state of the system can be split
as follows:
[ ]T
x(k) = x1T (k) x2T (k) · · · xm
T
(k) , (22)
∑ m
where xi ∈ ℝni and with i=1 ni = n. Therefore, the state dynamics of every subsystem is given by the following general
equation:
xi (k + 1) = Φi xi (k) + Γi ui (k), (23)
where [ ]
0I
Φi = 0 n0i −1 ∈ ℝni ×ni (24)

𝚪i = [0 0 · · · 0 Γi0 ]T ∈ ℝni ×m . (25)

Therefore, the global dynamic is expressed by


x(k + 1) = Φx(k) + Γui (k), (26)
where
Φ = diag [Φ1 , Φ2 , … , Φi , … , Φm ] ∈ ℝn×n
[ ]
𝚪T = 𝚪T1 𝚪T2 · · · 𝚪Tm ∈ ℝni ×m , 𝚪i ∈ ℝni ×m . (27)
Finally, from the MIMO controller canonical representation, the m sliding surfaces such that si = Ci xi , i = 1, … , m,
can be found in the same way that for scalar systems.
There exist a lot of research works where the discrete-time system dynamics is given in the control canonical form (14)
for DVSC process control where the switching surface is specified by arbitrary eigenvalue assignment. For example, in the
work of Chern et al,59 a discrete integral VSC system is proposed for controlling brushless DC motors using this canonical
representation.

Regular or reduced canonical form


One of the earliest and most used linear system canonical representations for sliding surfaces design is the regular
canonical representation,13,14 also known as reduced canonical form or as VSC design canonical form. In the regular rep-
resentation, the vector state x(k) is decomposed into two vectors x1 ∈ ℝn−m , x2 ∈ ℝm such that the system dynamics is
defined by the following difference equation:
x1 (k + 1) = Φ11 x1 (k) + Φ12 x2 (k) (28)
x2 (k + 1) = Φ21 x1 (k) + Φ22 x2 (k) + Γ2 u(k),
LUIS-DELGADO ET AL. 7

where u ∈ ℝm is the input of the system and with Φ11 ∈ ℝn−m×n−m , Φ12 ∈ ℝn−m×m , Φ21 ∈ ℝm×n−m , Φ11 ∈ ℝm×m , and
Γ2 ∈ ℝm×m such that Γ2 is not singular. The system dynamics can be seen as two reduced-order subsystems: the slow or
the lower subsystem {[ Φ21 Φ22 ], Γ2 } with u(k) inputs and state x2 (k) with the same order of u(k) and the faster or upper
subsystem {[ Φ11 Φ12 ], 0} with the vector x2 (k) as input and x1 (k) as the system state. From (28), it can be seen that the
upper subsystem is completely independent of the inputs u(k). If the switching surface is defined as follows:
s(k) = C1 x1 (k) + C2 x2 (k), C1 ∈ ℝn−m and C2 ∈ ℝm , (29)
then the ideal sliding phase dynamics is described as follows:
x1 (k + 1) = Φe x1 (k) (30)
x2 (k) = −C2−1 C1 x1 (k),
where
Φe = Φ11 − Φ12 C2−1 C1 . (31)
In (30), it is shown that, in the ideal SM, the system dynamics do not depend on the control law u(k). Therefore, the
design of the sliding surface is equivalent to the classical linear control design by state vector feedback. Therefore, the pair
(Φ11 , Φ12 ) is needed to be reachable or stabilizable. It can be shown that this condition holds if the pair (Φ, Γ) is reachable
or stabilizable, where [ ]
0
Γ= Γ . (32)
2

Then, an asymptotically stable sliding surface can be obtained defining of a matrix L = −C2−1 C1 , L ∈ ℝm×(n−m) such
that the eigenvalues of (31) can be set at some desired values. Once L is found, then s(k) can be calculated as follows:
s(k) = C2 [−L I]x(k), (33)
where C2 can be arbitrary set.
Normally, the system dynamics is not given in the reduced canonical form. Nevertheless, any controllable linear system
can be represented in a reduced canonical form using a similarity transformation. Considering Γ as two matrices Γ1 ∈
ℝ(n−m)×m and Γ2 ∈ ℝm×m such that Γ−12
≠ 0, then it is possible to obtain a regular representation (40) via the following
similarity transformation:
[ ]
In−m −Γ1 Γ−12
z(k) = Mx(k) = x(k). (34)
0 Γ−1
2
As a whole, a designing method to obtain a discrete-time switching surface s(k) = Cx(k) via (40) is as follows: (i) verify
that the system is controllable and reachable; (ii) apply the similarity transformation z = Mx to the system represented
by (34) to obtain the regular representation (28), where M is given by (34); (iii) define the following switching surface:
Sz = C1 z1 (k) + z2 (k); (35)
(iv) set the (n − m) eigenvalues or roots of the characteristic polynomial of the reduced-order system, ie,
z1 (k + 1) = [Φ11 − Φ12 C1 ]z1 (k); (36)
(v) find the feedback gain matrix C1 ∈ ℝm×n such that it makes system (36) stable via shifting poles; and (vi) apply the
inverse of the similarity transformation to C and estimate s(k) = Cx(k) via
C = [C1 Im ]M. (37)

3.1.2 Sliding surface design based on the Ackermann equation


As it was previously stated, the design of switching surfaces can be seen as a pole assignment method via state feedback
such that it is used in linear control theory. Consequently, some methods used for eigenvalue shifting can be applied to
sliding surface designing. Ackermann and Utkin18,60 and Utkin et al20 proposed a procedure for sliding surface designing
using Ackermann's formula used for pole assignment via state feedback. The two main advantages of this method are
as follows. Firstly, it is not necessary to obtain the sliding phase dynamics and, secondly, none specific representation is
needed.
8 LUIS-DELGADO ET AL.

In the work of Ackermann and Utkin,18 for a controllable scalar discrete-time linear system completely specified via
difference equation (40) and it is controlled via DVSC, it is proposed a sliding surface  = {Cx(k) = 0} with coefficients
CT ∈ ℝn obtained by the Ackermann's equation
C = eP1 (Φ), (38)
where
e = [0 0 · · · 1][Γ ΦΓ Φ2 Γ · · · Φn−2 Γ]−1 (39)
P1 (z) = (z − 𝜆1 )(z − 𝜆2 ) · · · (z − 𝜆n−1 )
= p1 + p2 𝜆 + p3 𝜆2 + · · · + 𝜆n−1 ,
such that  describes a hyperplane where the poles of the SM dynamic corresponds to 𝜆d = {𝜆1 , 𝜆2 , … , 𝜆n − 1 }.
An important disadvantage of this proposal is that it can not be directly applied to multi-input systems. However,
an extended design method of switching hyperplanes to multivariable LTI systems using an equation similar to the
Ackermann's formula was presented by Huang and Yeung,61 where the desired closed-loop eigenvalues of the sliding
hyperplanes are directly determined. Considering a discrete-time multi-input LTI system described as follows:
x(k + 1) = Φx(k) + Γu(k), (40)
where x ∈ ℝn , u(k) ∈ ℝp are the state and input vectors, respectively, and where Φ ∈ ℝn×n and Γ ∈ ℝn×p are known.
Matrix Γ is assumed to be full rank one and the pair (Φ, Γ) is also assumed to be controllable, ie, the rank of the controlla-
bility matrix Wc (Φ, Γ) ∈ ℝn×np is equal to n, where Wc (Φ, Γ) is defined in (15). Following the approaches applied to pole
assignment for multi-input LTI systems,57 it is needed to find the first n linear independent columns from Wc . Considering
the columns Γi ∈ ℝn of Γ, ie,
Γ = [Γ1 Γ2 · · · Γp ], (41)
then it is possible to rearrange Wc as follows:
[
Wc (Φ, Γ) = Γ1 ΦΓ1 · · · Φm1 −1 Γ1 ; Γ2 ΦΓ2 · · · Φm2 −1 Γ2 ; · · ·
]
· · · Γmi ΦΓmi · · · Φmp −1 Γp , (42)
∑p
where i=1 mi = n. This columns arrange of the controllable matrix implicitly considers that the system is divided into
p subsystems, where the ith subsystem is of mi order. Then, the arbitrary eigenvalue for the SM is grouped into p sets,
and for every ith subsystem, it is assigned (mi − 1) poles. Consequently, the complex poles must be assigned in pairs, so
the order of the ith subsystem must be high enough to include the pair of poles. In a numerical example of this review
work, this problem is well illustrated. Once the rearranged matrix is achieved, then the switching function s ∶ ℝn → ℝp
is defined as61
[ ]T
s = sT1 sT2 … sTm , si ∈ ℝn , (43)
such that s(k) = Cx(k), where C ∈ ℝp×n is given by
[ ]T
C = cT1 cT2 … cTm , ci ∈ ℝn (44)
and where every vector ci ∈ C is defined as follows:
ci = qTi 𝑓i (Φ) (45)

qTi = 𝛿i th row of Wc−1 (Φ, Γ) (46)


i
𝛿i th = m𝑗 (47)
𝑗=1


p
𝑓i (Φ) = 𝛼(𝜆) (48)
i=1

𝑓i (𝜆) = 𝜆mi −1 + 𝑓i,1 𝜆mi −2 + · · · + 𝑓 , (49)


then the eigenvalues of the ideal equivalent system are set in the roots of the polynomial 𝛼(𝜆).
A major problem with this methodology is the lack of precision in the case of MIMO or high-order systems. For contin-
uous time systems, Dorling and Zinober15,62 proposed a robust eigenvalues assignment procedure for the design of sliding
hyperplanes for continuous multivariable linear systems. However, for discrete-time systems, the authors have not found
LUIS-DELGADO ET AL. 9

any similar approach. Although the proposed method given in the aforementioned works15,62 was proposed for continu-
ous time system, we have decided including it because it was a relevant approach that can be useful for future research
applied to discrete-time systems. Consider the continuous time system dynamics expressed as follows:
.
x(t) = Ax(t) + Bu(t) + 𝑓 (t)u(t), (50)
where x ∈ ℝn , u(k) ∈ ℝm are the state and inputs of the system, respectively, and 𝑓 (t) ∈ ℝn×m are the disturbances of the
system. Matrices A ∈ ℝn×n and B ∈ ℝn×m are time invariant and completely known and it is also assumed that the pair
(A, B) is completely controllable. If the disturbance f (t) are matched, then the rank([B f (t)]) = rank(B) = m, therefore
the ideal SM is not disturbed. Representing the system dynamics in the regular canonical form (28) and as long as the
disturbances f (t) are not considered, then it has been shown that the ideal sliding dynamics is given by
.
x1 (t) = (A11 − A12 C1 )x1 (t), (51)
where the sliding surface is considered as (35). For multivariable systems, it is known that it is only needed n of the nm
degree of freedom for estimating the feedback matrix C1 and the remaining n(m − 1) of the degree of freedom can be used
for partially assigning the eigenvector of the equivalent system.62 Therefore, it is preferred to select the eigenvectors of the
equivalent system in the SM such that the eigenvalues associated to them can be robust to deviation to the parameters
of A. The robustness of the eigenvalues assignment method proposed in the works of Dorling and Zinober15,62 is in that
sense. Indeed, in these research works, a sensitivity metric ci for every eigenvalue 𝜆i of the closed-loop system equivalent
matrix (A11 − A12 C1 ) was suggested. The objectives of the design algorithm is to minimize the sensitivities of the spectrum
of the closed-loop equivalent matrix. Hence, a condition number 𝜅(V) estimate is defined as follows15 :
𝜅(V) = ||V||||V −1 ||,
where V = [v1 v2 · · · vn − m ]. Then, based on the robust pole assignment work developed in the work of Kautsky and
Nichols,63 Dorling and Zinober15 proposed the following iterative method for designing sliding surfaces.
1. Build a basis H(𝜆i ) for the null space S = Cx(t) = 0 as the columns of an n × m matrix W such that
[ ]
V
W = Ui
i

Si SiT = In−m .
2. Compute the vectors vi ∈ R (Si ) with the following restriction:
||vi || = 1, 1 ≤ i ≤ n − m.
3. Define V = [v1 v2 · · · vn − m ], (iv) define wi = SiT vi , 1 ≤ i ≤ n − m, and (v) find C1 such that C1 V = W.

3.1.3 Closed-equation algorithm for designing sliding surfaces via arbitrary


eigenvalue assignment
The proposal presented by Chen and Chang64 introduced a method for designing sliding surfaces for controlling contin-
uous systems in a very straightforward way. This proposal can be directly applied to discrete-time LTI systems as it is
summarized in Proposition 2.
Proposition 2. Given a controllable discrete system described as follows:
x(k + 1) = 𝚽x(k) + 𝚪u (k) , (52)
where x ∈ ℝn , u(k) ∈ ℝm . Considering that 𝚽 ∈ ℝn×n and 𝚪 ∈ ℝn×m are time invariant and completely known and the
rank(𝚪) = m. Choosing a sliding hyperplane of the following form:
s(k) = Cx(k)
(53)
s = [ s1 s2 … sm ],
with C ∈ ℝm×n , which fulfills that |C 𝚪| ≠ 0. Then, using the methodology proposed in the work of Chen and Chang,64
C can be computed as follows. (i) compute a pole-assignment feedback gain K ∈ ℝm×n such that the eigenvalues of the
hypothetical closed-loop system
x(k + 1) = [𝚽 − 𝚪K] x(k) (54)
10 LUIS-DELGADO ET AL.

are located in the desired positions, ie,


⎧ ⎫
⎪ ⎪
𝜆d = ⎨𝜆1 , 𝜆2 , … , 𝜆n−m , 𝜆, 𝜆, … , 𝜆⎬ , (55)
⎪ ⏟⏞⏞⏞⏟⏞⏞⏞⏟⎪
⎩ m𝜆 ⎭
where 𝜆 ≠ 𝜆i , i = 1, 2, … , n − m, and (ii) estimate the matrix C ∈ ℝm×n as follows:
C = K(𝚽 − 𝜆In )−1 . (56)

In this methodology, it is considered that 𝜆d ∩ Λ(A) = {∅}, so |A − 𝜆In | ≠ 0. It is also supposed that the matrix
[A − BK] is diagonalizable even though it has m repeated eigenvalues. Chen and Chang64 proved that, if the conditions
presented in the first point of the Proposition 2 are satisfied, then CB = Im and Λ{[In − B(CB)−1 C]}A = 𝜆d .
A modified approach based on this method was presented in the work of Draženović et al.21 In this work, the authors
proposed that the value of C could be obtained by resolving the following equations:
C(A − BK) = 0 (57)
CB = Im .
Hence, the algorithm is as follows: (i) estimate the gain matrix K ∈ ℝm×n as was explained in Proposition 2 and
(ii) calculate the matrix C ∈ ℝm×n as the solution of
C = [K Im ][A B]+ , (58)
where [·] + is the pseudoinverse of [·]. This algorithm can be applied directly to discrete-time linear systems, with the only
constraint that the characteristic polynomial of [A − BK] being strictly Schur.

3.2 Design of integral SM surfaces


Another extensive research line related to the specification of sliding surfaces applied to LTI discrete-time systems is the
integral SM control (ISMC) methodologies, which was introduced by Utkin and Shi.33 This robust method for designing
VSC eliminates the reaching phase and, consequently, the order of the SM is the same as the order of the controlled
system. Thus, the robustness of the closed-loop system can be guaranteed on the whole state-space. Two approaches
published by Abidi et al34,35 deal with the problem of designing sliding surface applied to discrete-time ISMC. This section
briefly presents the design methodology.
Consider the following disturbed discrete-time linear system:
x(k + 1) = Φx(k) + Γu(k) + 𝑓 (k), (59)
where x ∈ ℝn is the vector state, u ∈ ℝ is the inputs of the system, and 𝑓 ∈ ℝn are the disturbances that are considered
smooth and bounded. The parameters of the systems are represented by the matrices Φ ∈ ℝn×n and Γ ∈ ℝm such that
the pair (Φ, Γ) is controllable. The discrete time integral sliding surface is defined by
s(k) = Cx(k) + 𝜀(k) − Cx(0), (60)

𝜀(k) = 𝜀(k − 1) + Ex(k − 1), (61)


where C, E ∈ ℝm×n and 𝜀 ∈ ℝm . The term Cx(0) is included to eliminate the reaching phase. Abidi et al34,35 proved that it
is possible to find a matrix C such that |C 𝚪| ≠ 0 and there also exists a matrix K, where
E = −C(Φ − In + ΓC). (62)
Therefore, the closed-loop system dynamics is given by
x(k + 1) = (Φ − ΓK) x(k) + 𝜁 (k), (63)
where 𝜁 ∈ ℝn is O (T3 ) and where T is the sampling period. This result was obtained considering that the equivalent
control is described by [ ]
uequ (k) = −(C Γ)−1 C Φx (k) + 𝑓̂(k) , (64)
̂
where 𝑓 (k) represents the disturbance estimation defined as
𝑓̂(k) = x(k) − Φx (k − 1) − Γu (k − 1) . (65)
LUIS-DELGADO ET AL. 11

A method for designing the integral sliding surface using Abidi et al approach is as follows.34,35 (i) Find a matrix K
such that eigenvalues of (Φ − ΓK) are set inside the unit disk. Obviously, some desired eigenvalues can be chosen as
𝜆d = {𝜆i ∈ ℝ ∶ 𝜆i < 1, ∀i = 1, 2, … , n}. (ii) Select an arbitrary matrix C provided that the condition |C𝚪| ≠ 0 is fulfilled.
(iii) Calculate the matrix E via (62).
Other developments of switching surface design applied to ISMC can be found in the work of Castaños and Fridman.65

3.3 Sliding surface specification for LTI via Lyapunov stability approach
An earlier method for designing multivariable sliding surfaces based on the Lyapunov stability analysis was presented by
Su et al.28 This work introduced an elegant solution applied to multivariable LTI systems and also included a proposition
for multivariable nonlinear systems where nonlinear manifolds solutions are obtained. However, these proposals cannot
be directly applied to discrete-time systems because they are based on the solution of continuous time Lyapunov equation.
More recently, Kim and Bandyopadhyay66 presented another Lyapunov approach applied to discrete-time systems
considering a regular form representation. Considering that a stabilizable LTI discrete-time system given by
x(k + 1) = Φx(k) + Γu(k) (66)
can be asymptotically stabilizable via full state feedback u(k) = −K x(k) iff it exists P > 0, Q > 0 such that (Φ −
ΓK)T P(Φ − ΓK) − P < −Q holds. If the system in (66) is represented in the regular form as in (28) and it is controlled
via DVSC with the sliding surface
s(k) = Cx1 (k) + x2 (k), (67)
then it can be proved that the SM of the closed-loop system
x1 (k + 1) = (Φ11 − Φ12 C) x1 (k) (68)
x2 (k + 1) = −Cx1 (k),
where
C ≜ P22
−1 T
P12 (69)
is asymptotically stable if it is possible to find a matrix
[ ]
P P
P = P11 P12 (70)
21 22

such that, for some matrix Q ≥ 0, the following discrete linear Lyapunov equation is fulfilled, ie,
(Φ − ΓK)T P(Φ − ΓK) − P + Q < 0. (71)
This is an easier method for designing sliding surfaces. Nevertheless, it has the drawback that, if the system is not
represented in the regular form, then a previous similar transformation could be needed.

3.4 Optimal control approaches


Since quadratic optimal control applied to linear time systems yields to linear state feedback solutions, LQR approaches
are among the earlier considered methods for solving the problem of achieving a stable SM surface.13,14 More recently,
Kim et al24 used LMI numerical tools for solving this optimization problem. Indeed, LMI is currently a research area
where a wide number of proposals for specifying sliding surfaces have been continuously given.22,23,25,26,66-71
The design of optimal switching surfaces for DVSC process control is used in a wide range of applications. For example,
in the work of Jung et al,72 a proposal for controlling a DC-AC inverter via DVSC optimal sliding surface definition, where
a cost function can be expressed in the form given in (76), is used. The most relevant approaches that can be found in the
literature are as follows:
• Ricatti equation–based optimal control approaches;
• LMI methodology-based approaches;
• min-max optimal control approaches.
The next sections explain the main feature of them.
12 LUIS-DELGADO ET AL.

3.4.1 Ricatti equation–based optimal control approaches


For continuous time and in the sense of bringing some comparisons between continuous-time and discrete-time optimal
control approaches, Utkin and Yang13 applied the LQR optimal design approach for estimating sliding surfaces. They
proposed two performance quadratic indices, ie,

1
J1 = xT (t)Qx (t) dt (72)
2∫
ts

1 [ T ]
J2 = x (t)Qx(t) + uTequ (t)Ruequ (t) dt, (73)
2∫
ts

where Q ∈ ℝ n×n
and R ∈ ℝ m×m
are symmetric positive semidefinite matrices. For J1 index, the following sliding surface
definition was obtained:
[ ( T −1 T
) ]
s(x(t)) = − Q−1
22 A12 P + Q22 Q12 Im Mx(t), (74)
where P ∈ ℝ(m−n)×(m−n) is the solution of the following Ricatti equation:
( T
) ( −1 T T
)
0 = P A11 − A12 Q−1
22 Q12 + A11 − A12 Q22 Q12 P

22 A12 P + Q11 − Q12 Q22 Q12 .


T −1 T
− PA12 Q−1 (75)
In addition, Pieper73 and Kim et al24 proposed two methodologies to compute optimal switching surfaces for first-order
dynamics SM for perturbed continuous linear systems, where some Ricatti equation must be solved for the specification
of the sliding surface.
On the other hand, for linear discrete-time systems, where the equivalent ideal sliding dynamic was stated in (12) and
where the switching function is defined as s(k) = Cx(k) = 0 and for the following performance index J ∶ ℝn → ℝ:


J= x(k)T Q1 x(k), (76)
k=0

then the sliding surface proposed by Spurgeon30 and Edwards et al32 becomes
s(x) = ΓT Px (k), (77)
where P ∈ ℝn×n is a positive semidefinite (p.s.d.) matrix, which defines the following Lyapunov function:
V(k) = xT (k) Px (k). (78)
According to Lyapunov stability analysis, the closed-loop system is asymptotically stable if ΔV(k) = V(k + 1) − V(k) < 0.
Considering a symmetric definite matrix Q > 0 and using (12), then (78) yields to the following Ricatti equation:
ΔV(k) = ΦT P Φ − ΦT P Γ(ΓT P Γ)−1 ΓT P Φ − P = −Q, (79)
which can be also obtained from the index performance (76). Any matrix P that solves (79) guarantees an optimal SM
restricted by (79).

3.4.2 LMI approaches


The LMI is an efficient numerical tool that can be applied for obtaining the solution of optimization problems where
some convex objective function must be minimized subjected to some constraint functions.74 In related works,23-26,44,67,75-77
LMI-based approaches have been widely used for designing sliding surfaces for continuous time systems and also for
discrete-time systems (see other works22,27,68-71,78,79 ). In these research works, the switching surface specification is obtained
using LMI formulations, which allows that the sliding surface can be designed for perturbed systems with uncertain-
ties, even with mismatched ones.25 Some developments previously explained made use of LMI expressions for optimal
sliding surface specifications,24 Lyapunov-based sliding surface definitions,28 ISMC approach,34 and min-max controller
developments, as it will be seen later.31 Nevertheless, the research works included in this section are completely based on
LMI methods for obtaining sliding surfaces specifications. Finally, some research works for the design of switching sur-
faces applied to industrial processes have been proposed using LMI methods. For example, in the work of Lin et al,27 it
is proposed the use of DVSC for controlling a fluid power electrohydraulic system, where an optimal switching surface is
obtained using LMI approach.
LUIS-DELGADO ET AL. 13

According to sliding surfaces designing, LMI approaches have been applied to different scenarios such as perturbed and
time variant systems, eigenvalue assignment, and optimal designing. Now, the design problem for perturbed and time
variant discrete-time linear systems is presented.
Problem 2. Given a perturbed discrete-time linear system with time input and state delays

x(k + 1) = [Φ + ΔΦ(k)]x(k) + Φd x(k − d) + Γu(k) + Γh u (k − h) , (80)

where x(k) ∈ ℝn , u ∈ ℝm are the state and inputs of the system, respectively; and Φ ∈ ℝn , ∈ Γ(k) ∈ ℝn×m ,
Φd ∈ ℝn , Γd (k) ∈ ℝn×m are known constant matrices; and d ∈ ℤ and h ∈ ℤ are the known state and input delays.
The uncertainties are represented by the unknown matrix ΔΦ(k) ∈ ℝn , which are assumed to be norm bounded,
then a discrete-time surface should be found such that the sliding dynamic of the closed-loop controlled system be
asymptotically stable.
Similar problems has been analyzed in other works68-70,78,79 using LMI-based approaches. Consider the following regular
representation:

x1 (k + 1) = [Φ11 + ΔΦ11 (k)]x1 (k) + Φd11 x1 (k − d(k)) (81)


[ ]
+ Φ12 + ΔΦ12 (k) x2 (k)
[ ][ ]
Φ21 Φ22 x1 (k)
x2 (k + 1) = x2 (k) +
Φd21 z−d(k) Φd22 z−d(k)
+ Γu(k) + 𝑓 (k, x(k)),

where z−d(k) is a time variant delay d(k) and 𝑓 ∶ ℤ × ℝn → ℝm is a nonlinear function representing the external
disturbances and not modeled dynamics. It is also assumed that:

[ΔΦ11 (k) ΔΦ12 (k)] = E1 F1 (k)[H1 H3 ] (82)

ΔΦd11 = E2 F2 (k)H2 , (83)


where E1 , E2 , and Hi are known constant matrices and Fi (k) are unknown time varying matrices. Choosing the linear
sliding surface as the form (67), then the SM dynamics is described by

x1 (k + 1) = [Φ11 − Φ12 C + ΔΦ11 − ΔΦ12 C]x1 (k) (84)


+ [Φd11 + ΔΦd11 ]x1 (k − d(k)).

The solution of this problem given in the works of Nizar et al68,70 in terms of LMI will be detailed in the LMI solution of
the numerical example 2. Yan et al69 introduced one time-dependent sufficient condition for the existence of the sliding
surface in LMI terms, giving one solution for specifying C ∈ ℝm×(n−m) and proving that the following sliding surface
definition:
s(k) = KL−1
1 x1 (k) + x2 (k) = 0 (85)
guarantees that the sliding dynamic (84) is asymptotically stable if, for certain positive number 𝜅, there exists some positive
scalars 𝛿 1 , 𝛿 2 > 0 and the (n − m) × (n − m) matrices L1 > 0, L2 , L3 , U1 > 0, U2 > 0, W ̄ 1, W
̄ 2, W
̄ 3 and the matrix
K ∈ ℝm×(n−m) such that the following LMIs is fulfilled:
⎡ Θ1 Θ2 0 0 L1 𝜗 LT2 LT2 ⎤
⎢ ⎥
⎢ ∗ Θ3 𝜍 0 0 0 L3 T
LT3 ⎥
⎢ ∗ ∗ −U1 U1 H2T 0 0 0 0 ⎥
⎢ ⎥
⎢ ∗ ∗ ∗ −𝛿2 I 0 0 0 0 ⎥ < 0, (86)
⎢ ∗ ∗ ∗ ∗ 𝜒 0 0 0 ⎥
⎢ ∗ ∗ ∗ ∗ ∗ −𝛿1 I 0 0 ⎥
⎢ ∗ ∗ ∗ ∗ ∗ ∗ −L1 0 ⎥
⎢ ∗ ∗ −d−1 ⎥
⎣ ∗ ∗ ∗ ∗ ∗ M
U 2 ⎦
[ ̄ ̄ ]
W1 W2 0
̄ 3 𝜅Φd U1 ≥ 0,
∗ W (87)
∗ ∗ U1
14 LUIS-DELGADO ET AL.

where symbol ∗ represent a term that is induced by symmetry and where


𝜗 = L1 H1T − K T H3T
𝜍 = (1 − 𝜅)Φd U1
𝜒 = −(dM − dm + 1)−1 U1
̄1
Θ1 = L2 + LT2 + dM W
( T )
̄2
Θ2 = L3 + L1 Φ + 𝜅ΦTd − I − K T ΦT12 − LT2 + dM W
Θ3 = −L3 − LT + dM W ̄ 3 + Σ2 𝛿i Ei ET ,
3 i=1 i

and where it is considered that the initial values of the system are given by x1 (k) = 𝜑1 (k), k = −dM , − dM + 1, … , 0.
A more recent research work by Tapia et al22 deals with a LMI-based design method for arbitrary eigenvalues assign-
ment. This research work includes one proposal for full order and another one for reduced order, both of them applied
to continuous and discrete-time systems. For the reduced-order discrete-time case, it is proved that the SM dynamics for
the sliding surface s = Cx has the pole assignment (A − BK) if the LMI in the following conditions hold:
0 > M + M T + (A − BK)T P(A − BK) − P, (88)
−𝜖 ≺ −M(A − BK) ≺ 𝜖,
where 𝜖 ∈ ℝ is a small arbitrary positive number, P = [P1 P2 ] is a p.s.d. matrix with P1 ∈ ℝn×m , P1 ∈ ℝn×(n−m) , and
M ∈ ℝn×n as decision variables, and a sliding subspace matrix given by C = P1+ M.

3.4.3 Min-max controller approach


Min-max controller approaches80 are another optimal methodologies applied to perturbed linear systems, which are
based on minimizing the control law effort supposing the worst disturbances conditions. Let us consider a disturbed
discrete-time linear system given as follows:
x(k + 1) = Φx(k) + Γ[u(k) + 𝜉(k)] (89)
𝑦(k) = Cx(k),
where x ∈ ℝn , u ∈ ℝm , and 𝑦 ∈ ℝp , where m < p < n. Assume that Γ and C are of full rank and (Φ, Γ, C) are
minimal. The matched uncertainties 𝜉(k) are supposed to belong to a balanced set, ie, if 𝜉(k) ∈  , then −𝜉(k) ∈  are
assumed to be unknown but bounded with known upper-lower limits. Define the Lyapunov function V(x) = x(k)T Px(k),
where P ∈ ℝn×n is a symmetric positive definite matrix. As stated in the work o Sharav-Schapiro et al,80 the optimal state
feedback control law that minimizes ΔV(k) = V(k + 1) − V(k) with respect to u(k) is given by
usmm (k) = −(ΓT P Γ)−1 ΓT P Φx(k), (90)
where the maximum bound of 𝜉(k) was considered. This control law guarantees the minimal u(k) for the maximal dis-
turbance 𝜉(k), hence, it is referred as Lyapunov min-max control law.26,80 Using the Ricatti equation previously defined in
(79), then
ΔV(k) = −xT (k)Qx(k) + 𝜉 T (k) ΓT P Γ𝜉(k). (91)
In absence of uncertainties, if Q > 0, then the system is asymptotically stable. It can be verified that Q ≡ P − ΦTc PΦc ,
where
Φc = Φ − Γ(ΓT P Γ)−1 ΓT PG. (92)
For the output feedback min-max controller,26 the p.s.d. matrix P must be chosen so that the constraint
ΓT PΦ = FC (93)
holds for some F ∈ ℝm×p . Thus, from the state feedback control law equation in (90), the following output feedback
control law can be derived:
uomm (k) = −(CΦ−1 Γ)−1 𝑦(k). (94)

A connection between output min-max controller and output DVSC was analyzed by Lai et al.31 Considering that, for
a switching surface s(k) = Fx(k), the ideal SM is defined as follows:
x(k + 1) = [In − Γ( F Γ)−1 F]Φx(k), (95)
LUIS-DELGADO ET AL. 15

then, replacing the state min-max control law (90) in the nominal system equation, ie, 𝜉(k) = 0, the closed-loop dynamics
is given by
[ ( )−1 ]
x(k + 1) = In − Γ ΓT P Γ ΓT P Φx(k). (96)
Comparing both (95) and (96) equations, then the control law is directly obtained as follows:
ssmm (k) = ΓT Px(k). (97)
It is important to note that it is exactly the same result started by Spurgeon.30 It is not difficult to show that, for the output
feedback min-max controller, the sliding surface is as follows:
somm (k) = C Γ−1 x(k). (98)

3.5 Flexible sliding surface design considering nonideal SM dynamic approach


In 2016, Luis-Delgado et al39 presented a novel approach that facilitates the task of designing sliding surfaces applied to
MIMO LTI systems, where the switching surface is defined through a simple equation that allows both eigenvalues place-
ment design or optimal methods without any coordinate system representation transformation. This approach guarantees
high accuracy of the system dynamics, either for arbitrary eigenvalues assignment or for optimal switching surface defi-
nitions and, in addition, make use of a design parameter for the adjustment of the nonideal SM dynamics. Given a MIMO
LTI discrete-time system with dynamics described in (9) and supposing as follows.
1. A gain matrix K ∈ ℝm×n that stabilizes the system via eigenvalues shifting such that the roots of the characteristic
polynomial of the matrix
XCL (𝜆) = det (𝜆In − Φ + 𝚪K) = 0 (99)
is located at the arbitrary eigenvalues
⎧ ⎫
⎪ ⎪
𝜆d = ⎨𝜆1 , 𝜆2 , … , 𝜆n−m , 0, 0, … , 0⎬ , (100)
⎪ ⏟⏞⏞⏟⏞⏞⏟⎪
⎩ m ⎭
where, if i ≠ j, then 𝜆i ≠ 𝜆j .
2. An arbitrary nonsingular design matrix E ∈ ℝm×m .
Then, if the switching function gain matrix is estimated as follows:
C = EKΦ−1 , (101)
then s(k) = Cx(k) makes the ideal SM of the controlled system asymptotically stable with eigenvalue located at 𝜆d , where
E = C𝚪 holds when K fulfills (99) and (100) is fulfilled. The sliding surface definition (101) brings a simple way to
configure the SM dynamics by mean of the arbitrary eigenvalues assignment.
The arbitrary matrix E = C𝚪 guarantees that, for an ideal SM, the invariance property14 holds. However, due to that for
discrete-time system, the ideal SM is not achieved and a quasi-SM appears,4 where the state vector oscillates in a region
around s(k) = 0. That implies that control laws must be designed to make sliding regions as thin as possible. Hence,
calculating the nonideal SM dynamics is made as follows:
[ ]
s(k + 1) = E K x(k) + u(k) , (102)
then it is obvious that s(k + 1) depends on E. Indeed, every si ∈ s(k) can be adjusted through a linear combination of the
components of the vector u(k). Consequently, E should be considered when the control laws are specified. In Luis et al
methodology,39 it is also shown that, for a multi-input LTI discrete-time system with dynamics described in (9) and given
a state feedback gain matrix K ∈ ℝm×n such that for the Lyapunov function
V(k) = xT (k)Px(k), (103)
the following linear Lyapunov equation holds:
[Φ − ΓK]T P[Φ − ΓK] − P = −Q, (104)
16 LUIS-DELGADO ET AL.

then the switching function s(k) = Cx(k), where C is given in (101) achieves an asymptotically stable SM dynamics.
Indeed, if the state feedback gain matrix K ∈ ℝm×n is achieved such that the performance index given in (76) is minimized,
then the switching function s(k) = Cx(k), where C that is given in (101) produces an optimal asymptotically stable SM
dynamics. Therefore, if the feedback gain K is calculated via LQR, then an optimal SM dynamics for the switching surface
obtained through (101) is achieved. It is important to note that, if the state feedback gain matrix K does not hold the
conditions given in (99) and (100), then E ≠ C𝚪. However, it can be shown that a desired C𝚪 = D can be achieved if E
is defined as follows:
E = D(KΦ−1 Γ)−1 . (105)

4 NONLINEAR D ISCRET E-TIME SYSTEMS SM HYPERPLANE DESIGN

A very extensive number of design methods have been explained so far. All of them were developed to be applied to linear
systems. In contrast, the number of research works applied to nonlinear systems is considerably lower. In order to give a
solution to this problem, some research works have proposed to obtain a linear approximation of the original nonlinear
system, then the switching surface is obtained considering the linear approximation. However, this methodology can be
applied to a few class of nonlinear system. One of the earliest works related with nonlinear sliding surface design was
presented by Sira-Ramírez,41 where it is proposed to obtain a linearized equivalent representation via state feedback.
Specifically, the linear representation is Brunovsky's canonical form and then a linear sliding surface is designed in the
linearized coordinates, exhibiting stable dynamics. Other approaches made use of linear approximation applied to process
control, such as the thermal mixing tank considered in the work of Camacho and Smith,50 where the use of a switching
surface as it is given in (158) is proposed so that the nonlinear system is approximately modeled via first-order plus
dead time (FOPDT) In another work, Sira-Ramírez proposed a normal representation for the design of discrete VSC
for nonlinear systems,3 although it was not presented a sliding surface design procedure. De-Carlos et al presented a
method for designing SM surfaces using regular canonical representation42,81 and Spurgeon and Davies43 made use of
noncanonical transformation. Beside this, Su et al28,44 proposed a Lyapunov approach for sliding surface design. More
recently, Gaffari and Yasdanpanah45 proposed a method for computing a nonlinear stable SM surfaces applied to nonaffine
continuous time systems and Dong and Shi47 introduced an algorithm for designing optimal sliding surface for nonlinear
discrete-time systems using the nonlinear two-point boundary problem approach. Finally, some approaches intended to
apply DVSC with SM to discrete-time nonlinear systems based on T-S fuzzy models46,48 had proposed linear methods such
that LMI.
The general nonlinear sliding surface problem can be stated as follows.
Problem 3. Consider a nonlinear discrete-time system whose dynamics is defined as follows:
x(k + 1) = 𝑓 (x(k), u(k), k) (106)
𝑦(k) = h (x(k), u(k), k) ,
where x ∈ X ⊂ ℝn ,
u∈U⊂ ℝm ,
and 𝑦 ∈ U ⊂ ℝp
are the state, input, and output vector, respectivel; and f ∶ X × U →
X and g ∶ X × U → Y are supposed to be analytical functions. Then, given a control law u(k) such that it achieves a
SM on the surface
 = {x ∈ X | s(x(k)) = 0}, (107)
then the switching function s ∶ X → Rm must be set to make the closed-loop system asymptotically stable.
Basically, the problem of finding a sliding surface for nonlinear system is the same as the linear case. However, the
methodology is not the same because some concepts applied to linear systems, such as arbitrary eigenvalue assignment,
cannot be used for solving nonlinear problems. Before starting the explanation of the design methods of sliding sur-
face applied to nonlinear system, let us find a general state dynamics. Consider a nonlinear discrete-time system whose
dynamics is defined in (106). Then, the equivalent control law uequ (k) can be found so that
s(k + 1) = s(k) = 0, (108)
where
( ) ( ( ))
s(k + 1) = [s ◦ 𝑓 ] x(k), uequ (k), k = s 𝑓 x(k), uequ (k), k . (109)
LUIS-DELGADO ET AL. 17

This equation yields to the ideal SM described as follows:

x(k + 1) = 𝑓 (x(k), uequ (k), k), k ∈ ℤ+ (110)


s (x(k)) = 0, ∀k.

For a linear switching surface s(k) = Cx(k) and from (106), then

C𝑓 (x(k), uequ (k), k) = 0, (111)

therefore
uequ (k) = 𝑓 −1 (x(k), u(k), k). (112)

In the case of affine systems, ie, systems whose dynamics exhibit linear behaviors in relation to the inputs and nonlinear
ones with respect to the states (106), it can be written as follows:

x(k + 1) = 𝑓 (x(k), k) + Γu(k). (113)

Thus, the sliding surface dynamic is defined as follows:

s(k) = C𝑓 (x(k), k) + C Γu(k), (114)

and the equivalent control law is given by

uequ (k) = −(C Γ)−1 C𝑓 (x(k), k). (115)

Therefore, the closed-loop ideal SM is described as follows:


[ ]
x(k + 1) = In − Γ(C Γ)−1 C 𝑓 (x(k), k). (116)

In general, the methods for designing nonlinear surfaces are based on obtaining a suitable canonical representation to
find an easier mathematical model that could help the design process.3,16,17,28,41-43 The following sections introduce some
canonical representations applied to nonlinear system DVSC, which are briefly explained.
The approaches applied to nonlinear system that have been considered in this survey and will be described in the next
sections are the following ones.
• nonlinear canonical representations approaches;
• Lyapunov stability analysis-based approach;
• flexible sliding surface design considering nonideal SM dynamic approach applied to nonlinear MIMO systems.

4.1 Nonlinear canonical representations approaches


Due to the complexity of the nonlinear systems dynamics, it is very useful to find some canonical representation that
allows an easier SM analysis to the designer. The following sections explain some typical nonlinear canonical repre-
sentations used for designing sliding surfaces. Some approaches use linearizing canonical representations, for example,
Brunovsky's canonical form,28,41,82,83 such that some linear sliding surface is designed in the equivalent linear coordinate
systems, which actually is a nonlinear sliding surface in the original state space coordinates. Canonical represen-
tation approaches are widely used in VSC and DVSC control of nonlinear processes. For instance, in the work of
Sira-RamÍrez et al,82 the canonical representation of the system dynamics is used for the design of switching surfaces for
SM control of chemical process.
The most used canonical representations of nonlinear systems dynamics for the design of sliding surface are the
following:
• reduced or regular canonical representation;
• controller and Brunovsky's canonical forms;
• normal canonical representation.
The following sections detail how this canonical representation are used for the specification of sliding switching surfaces.
18 LUIS-DELGADO ET AL.

4.1.1 Reduced or regular canonical representation


For nonlinear discrete-time systems, a reduced canonical representation is as follows:
x1 (k + 1) = 𝑓1 (x(k), k) (117)
x2 (k + 1) = 𝑓2 (x(k), k) + Γ(x)u(k), (118)
where x1 ∈ X1 ⊂ ℝ n−m
, x2 ∈ X2 ⊂ ℝ , u ∈ U ⊂ ℝ , f1 ∶ X1 → X1 , and f2 ∶ X2 → X1 , with Γ(x(k)) ∈ ℝ
m m n×m
not being

singular and x = [x1T (k) x2T (k)]T .


Supposing a controlled nonlinear system such that the SM on s(k) = 0 is achieved. If s(k) is linear in x2 (k), ie,
s(k) = s1 (x1 (k)) + C2 x2 (k), (119)
then s1 ∶ ℝ n−m
→ ℝ must be designed such that the SM is asymptotically stable. From the equivalent control approach
m
−1
x2 (k) = −C2 s1 (x1 (k)), therefore the reduced order ideal SM is given by
( )
x1 (k + 1) = 𝑓1 x1 (k), −C2−1 s1 (x1 (k)), k . (120)
For the case of a linear surface s1 (x1 (k)) = C1 x1 (k), then, using equivalent control approach, x2 (k) = −C2−1 C1 x1 (k), thus
the reduced system SM dynamics is given by
( )
x1 (k + 1) = 𝑓1 x1 (k), −C2−1 C1 x1 (k), k , (121)
where C2 ∈ ℝm×m is assumed to be nonsingular. Supposing x(k) = 0 to be an equilibrium state, then the problem of
designing the sliding surface for nonlinear systems is restricted to finding C1 , C2 such that the reduced SM dynamics (121)
must be asymptotically stable. It is likely that the design of sliding surfaces for linear SM dynamics behavior yields to
nonlinear surfaces.16,28,42
In general, the second Lyapunov criterion can be used for analyzing the stability of the SM. Hence, given a Lyapunov
function V(k) such that
ΔV(k) = V(k + 1) − V(k) < 0, (122)
then, if the following verification holds:
( ( ))
V 𝑓1 x1 (k), −C2−1 C1 x1 (k), k − V(x(k)) < 0, (123)
the ideal SM dynamic (120) is asymptotically stable.
If the system dynamics is given in the general form (106), then a nonlinear coordinate transformation z(k) = g(x(k))
must be found, which usually is an involved task. Further information about regular transformation applied to VSC design
can be found in other works.16,17,42

4.1.2 Controller and Brunovsky's canonical forms


For nonlinear systems with m input signals, a system representation can be obtained as m partitioned subsystems, where
each subsystem is given in the controller canonical representation such that84
[ ]
xT = x1T x2T · · · xm
T
, (124)
∑m
where xi ∈ ℝ i such that i=1 ni = n. Every subsystem dynamics is given by the following equation:
n

xi (k + 1) = Φi xi (k) + 𝛂i (x(k)) + Γi (x(k)) ui (k) , (125)


where [ ]
0 Ini −1
Φi = ∈ ℝni ×ni
0 0
𝛂i (x(k)) = [0 0 · · · 0 𝛼i0 (xi (k))]T ∈ ℝni
T (126)
⎡0 ··· 0 ··· 0⎤
⎢⋮ ··· ⋮ ··· ⋮⎥
𝚪i (x(k)) = ⎢ 0 0⎥ ∈ ℝ .
ni ×m
··· 0 ···
⎢0 · · · Γi0 (xi (k)) ··· ⎥
0⎦

Hence, the total dynamics can be expressed as follows:
x(k + 1) = Φx(k) + 𝛂(x(k)) + Γ(x(k))u(k), (127)
LUIS-DELGADO ET AL. 19

where
Φ = diag [Φ1 , Φ2 , … , Φi , … , Φm ] ∈ ℝn×n
[ ]
𝛂(x(k))T = 𝛂1 (x(k)) 𝛂2 (x(k)) … 𝛂m (x(k)) ∈ ℝni (128)
[ ]
𝚪(x(k))T = 𝚪1 (x(k)) 𝚪2 (x(k)) … 𝚪m (x(k)) ∈ ℝni ×m .
If a canonical controller representation is achieved, then m decoupled subsystems are obtained and, therefore, m slid-
ing surfaces si (k) = Ci xi (k) can be estimated using some of the linear switching surface design approaches previously
explained.
The subsystem dynamics representation given in (125) is also known as Brunovsky's canonical representation. For
scalar systems, it is given as follows:
xi (k + 1) = xi+1 , i = 1, 2, … , n − 1 (129)
xn (k + 1) = 𝑓 (x) + g(x)u (k) .
From (129), it is clear that, if the switching surface is as given in (5), then a suitable sliding surface specification can be
obtained applying any linear design approach. In the work of Lee et al,85 it is proposed some conditions for the existence
of a diffeomorphism Ψ(x(k), u(k)) for transforming from a nonlinear system dynamics representation of the general form
(106) into a Brunovsky's canonical one. Brunovsky's canonical form is frequently used for the specification of switching
surfaces in DVSC and VSC control of nonlinear systems. Some examples can be seen in related works.28,41,44,56,86

4.1.3 Normal canonical representation


Another canonical representation used for nonlinear MIMO systems is the normal canonical representation with zero
dynamics.3,49,84,87 For understanding this canonical representation, it is important to know the meaning of the relative
degree parameter.
Definition 3. Given a MIMO nonlinear system with dynamics described as in (106). If ri is the smallest integer where
r −1
Lg L𝑓i hi (x) ≠ [0 0 · · · 0], then the pair ( f, h) is said to have a relative degree (or relative order) ri of the output yi (x(k))
with respect to the input signals u(k), where Lf h(x) is the Lie derivative of h(x) in the direction of f(x) and where
r −2 r −1
r −1 𝜕(L𝑓i hi (x)) r −1 𝜕(L𝑓i hi (x))
L𝑓i hi (x) = 𝜕x
𝑓 (x) and Lg L𝑓i hi (x) = 𝜕x
g(x).
The definition of relative degree given in Definition 3 is stated for every single output. Hence, the total system relative
degree r is defined as follows:
∑m
r= ri . (130)
i=1

For discrete-time systems, the relative degree ri is the smallest number of sampled periods after which a change in u(k)
affects the output yi (k). The following proposition describes the procedure for obtaining a discrete-time MIMO normal
form from a general system representation, as given in (106).
Proposition 3. Consider a completely known nonlinear process, which is described by (106) with relative orders ri .
Choosing the nonlinear invertible coordinate transformation zi = Φi (x) with
Φi (x) = col[𝜙1 (x) 𝜙2 (x) · · · 𝜙n (x)],
where 𝜙j (x) = h ◦ f j − 1 (x) for 1 ≤ j ≤ ri and qr + j (x) = 𝜑j for 1 ≤ j ≤ n − r with 𝜑j arbitrary and independent
functions from the previous ri ones. The transformed system over z 0 = Φ(x 0 ) is
zi (k + 1) = zi+1 (k), for 1 ≤ i < r1−1 (131)

zri (k + 1) = [h ◦ 𝑓 i−1 ](Φ−1 (x(k)), u(k)), i = 1, … , m

zr+𝑗 (k + 1) = q𝑗 (z) for.1 ≤ 𝑗 ≤ n − r



i
𝑦i (k) = zl (k), l = r𝑗 .
𝑗=1
20 LUIS-DELGADO ET AL.

The zero dynamics are defined as the reduced submodel dynamic that results from choosing the input signals ui (k) and
the initial state x(0) such that yi (k) = hi (x) = 0, ∀k ≥ 0 and i = 1, … , m. Therefore, the zero dynamic corresponds
to zr − 1 (k + 1), … , zn (k + 1) with coordinates z1 , … , zr . With 𝜉 = [z1 , … , zr ]T and 𝜂 = [zr + 1 , … , zn ]T , then the zero
dynamic can be written as
𝜂(k + 1) = q(𝜉, 𝜂) = q(0, 𝜂). (132)

If a control law u(k) is chosen such that

[h ◦ 𝑓 i−1 ](Φ−1 (x(k)), u(k)) = 0, (133)

then the system state remains over the surface 𝜉 = 0. Therefore, the sliding surface corresponds to the zero dynamics of
the system, where Equation (133) describes the sliding dynamics on s(x) = 0 (see the work of Sira-Ramírez3 ). The normal
form is widely used for the design of nonlinear VSC, for example, for the control of chemical processes.88,89

4.2 Flexible sliding surface design considering nonideal SM dynamic approach applied
to nonlinear MIMO systems
In a recent research work, Luis-Delgado et al40 presented a switching surface equation for a class of MIMO nonlinear
systems, where an explicit equation similar to the one presented in another work of Luis-Delgado et al,39 which makes
use of an implicit linearizing process via Taylor expansion, allows the use of linear procedures for the design of switch-
ing surfaces and the forward Euler method to obtain a discrete time dynamics representation. In a different work of
Luis-Delgado et al,40 it is shown that, for a MIMO nonlinear continuous time system (·) with dynamics described as
follows:
.
x(t) = f (x (t) , u(t)) , (134)

where x ∈ X ⊂ ℝn is the state vector and f ∶ ℝn+m → ℝn defined as follows:

f(x) = [𝑓1 (x, u) 𝑓2 (x, u) · · · 𝑓n (x, u)]T , (135)

where every 𝑓i ∶ ℝn+m → ℝ is smooth nonlinear functions, then a state feedback gain matrix K ∈ ℝm×n such that the
following system dynamic is asymptotically stable:
̂ − 𝚪K]x(k),
x(k + 1) = [𝚽 ̂ (136)
̂ and 𝚪̂ are defined as follows:
where 𝚽
̂ = [In + TJA ]
𝚽 (137)
𝚪̂ = TJB ,

where In is the nth-order identity matrix, T is the sample time, and JA and JB are the Jacobian matrix of f(x) with respect
to x(t) and u(t), respectively, such that
𝜕𝑓 || 𝜕𝑓 ||
JA (x) = | , JB (x) = ,
𝜕x |x=xe ,u=ue 𝜕u ||x=xe ,u=ue

and where xe ∈ X is an equilibrium state. Therefore, if (·) is controlled via a DVSC, then the switching function s(k) =
Cx(k) make the discrete-time ideal SM of the controlled system asymptotically stable around xe ∈ X, if C is estimated as

C = EK[In + TJA (x)]−1 , (138)

where E ∈ ℝm×m is a nonsingular arbitrary design matrix and [In + TJA (x)] is also a nonsingular matrix. Additionally,
the work presented by Luis-Delgado et al40 shows that the sliding surface s(k) = Cx(k), where C ∈ ℝm×n is defined
in (138), guarantees the asymptotic stability of any MIMO nonlinear system represented as in (134) and where a state
feedback gain matrix K is needed. Hence, the problem of designing the sliding surface s(k) = Cx(k) is reduced to the
specification of matrix C so that the characteristic polynomial of 𝚽 ̂ − 𝚪K]
̂ CL = [𝚽 ̂ is strictly Schur.55 Additionally, the gain
matrix K can be either estimated via arbitrary eigenvalues assignment approaches or via optimal methods such as LQR
or linear quadratic Gaussian. This can be considered as an advantage of the proposed methodology because the designers
LUIS-DELGADO ET AL. 21

can obtain several Ks and then the evaluation of the SM performance can be obtained for every K. For the pole placement
methodology, it is worth mentioning that m eigenvalues of 𝚽̂ CL must be set to zero because the equivalent SM always has
90
m poles at the origin. The other n − m eigenvalues must be set inside the unit circle.

4.3 Lyapunov stability analysis approach


Another method for constructing a switching surface is based on the Lyapunov stability theory.44 Let us consider an affine
continuous nonlinear system as

x(t) = 𝑓(x(t)) + g(x)u. (139)

In order to apply the second or direct Lyapunov method, the next requirements must be satisfied.
1. A feasible feedback function u(k) = k(x) must be found such that the controlled system is asymptotically stable.
. .
2. Some Lyapunov function V(x) should exist such that V(x) ≤ 0 and V(x) ≠ 0 ∀x ≠ 0.
Su et al44 proved that the sliding surface
( )T
𝜕V(x)
s(x) = gT (x) = 0. (140)
𝜕x

make system (139) asymptotically stable in the SM. Therefore, the main difficulty of designing sliding surfaces with this
method proposed in the work of Su et al44 consists in finding the Lyapunov V(x) because it is usually very difficult to
specify Lyapunov function for complex nonlinear systems.
Some research works related to process control made use of Lyapunov approaches for the design of switching surfaces.
For example, in the work of Aliakbari et al,91 a fault-tolerant control proposal for heat recovery steam generator boiler
in combined cycle power plants is implemented with SM switching surfaces, which are estimated using a Lyapunov
approach.

5 CHATTERING REDUCTION

It has already been commented that one of the main differences of discrete-time variable radius controllers with those
of continuous time radiates in the existence of the nonideal SM, which corresponds with the oscillations or chattering
around the sliding planes. In39 it is shown that the amplitude around the switching surface is described by:

s(k + 1) = (1 − qT)s(k)𝜖T signs[s(k)]. (141)

In the sliding surface design techniques, the reduction of the SM can only be implemented when designing the control
laws, since the parameters of the sliding surface are estimated considering that the controlled system presents an ideal
dynamic in the mode sliding, that is, it is not possible to reduce the oscillations by specifying the sliding surface because
it does not correspond to the ideal dynamic. Therefore, the schemes for reduction the chattering phenomenon consist
on smoothing the control law signals (see16,92,93 ). A general scheme that is commonly applied to eliminate chattering
oscillation is given in Figure 2, where it can be seen three main blocks for the implementation of a DVSC controller:
i) the switching surface module, ii) the control law estimation and iii) the anti-chattering scheme.
However, the research work presented in39,40 introduces a design parameter of the switching surface definition that can
be adjusted for minimizing the amplitude of the oscillations. In,39 the authors show that if the sliding surface is estimated
using (101) then, the nonideal SM dynamics can be described as follows:
[ ]
s(k + 1) = E K x(k) + u(k) . (142)

Hence, it is clear that the amplitude of s(k + 1) depends directly on E. Indeed, E does not affect the ideal SM dynamics
(see39 ) then the control vector u(k) can be estimated and, after that, the parameter matrix E is computed for reducing the
chattering amplitude without adding any scheme.
22 LUIS-DELGADO ET AL.

FIGURE 2 General scheme of a closed-loop system controlled via discrete-time variable structure control (DVSC) with antichattering
[Colour figure can be viewed at wileyonlinelibrary.com]

6 ILLUSTRAT IVE EXAMPLES

To illustrate the previously explained design procedures, three examples of multivariable processes are considered. The
numerical calculation is made using the computing environment MATLAB®. Additionally, two examples of common
control process with some references to previous research works are presented.
Example 1 (LTI MIMO process).
Let us consider the same fifth-order linearized model of the aircraft used in the works of Draženović et al21 and
Tapia et al,22 where the state variables are as follows: x1 = height error relative to ground or guidance aid, x2 = for-
ward speed (m∕s), x3 = pitch angle (degrees), x4 = rate of change of pitch angle (degree∕m), and x5 = vertical speed
(m∕s). The inputs correspond to u1 = spoiler angle (degrees), u2 = forward acceleration due to engine thrust (m∕s),
and u3 = elevator angle (degrees). The continuous time dynamics can be seen in the works of Draženović et al21 and
Tapia et al.22 Considering a sample period T = 0.1 s, then the discrete-time dynamics are governed by the difference
equations of the form (9), where

⎡1 0.0014 0.1132 0.0005 −0.0970 ⎤


⎢0 0.9935 −0.0171 −0.0005 0.0068 ⎥
Φ = ⎢0 0.0003 1 0.0957 −0.0048 ⎥ , (143)
⎢0 0.0059 0 0.9131 −0.0940 ⎥
⎢0 −0.0278 0.0002 0.0977 0.9361 ⎥⎦

⎡ −0.0077 0 0.0003 ⎤
⎢ −0.0115 0.0997 0 ⎥
Γ = ⎢ 0.0212 0 −0.0080 ⎥ . (144)
⎢ 0.4152 0.0003 −0.1581 ⎥
⎢ 0.1749 −0.0014 −0.0154 ⎥
⎣ ⎦

The eigenvalues of the discretized system matrix Φ are Λ(Φ) = {1, 0.9238 ± 0.0950i, 0.9976 ± 0.0184i}. Then, a sliding
surface must be designed such that the eigenvalues of the ideal SM system matrix Φequ should be shifted to 𝜆d =
{0.7 ± 0.7i}.

Solution 1 (Controller canonical approach).


From the controllable matrix (15), an nth-order linear independent column matrix W is obtained such that

W = [Γ1 ΦΓ1 Γ2 ΦΓ2 Γ3 ], (145)


LUIS-DELGADO ET AL. 23

̂ = T −1 ΦT and Γ̂ = T −1 Γ is
and from (145), the transformation matrix T is estimated57 and the transformed system Φ
obtained as follows:
⎡ 0 1 0 0 0 ⎤
⎢ −0.8862 1.8754 −0.0006 0.0004 0.0033 ⎥
̂ =⎢
Φ 0 0 0 1 0 ⎥
⎢ 10.2086 −3.1786 −6.8476 3.9564 66.5266 ⎥
⎢ −0.2103 −0.9891 ⎥⎦
⎣ 0 0.1163 0
⎡ 0 ⎤ ⎡0 0 0 ⎤
⎢ Γ̂ 1 ⎥ ⎢ 1 0 −0.3586 ⎥
Γ̂ = ⎢ 0 ⎥ = ⎢ 0 0 0 ⎥.
⎢ Γ̂ ⎥ ⎢ 0 1 −33.5963 ⎥
⎢ 2⎥ ⎢0 0 ⎥
⎣ Γ̂ 3 ⎦ ⎣ 1 ⎦

With this equivalent system, matrix C can be obtained as follows: from the equivalent control theory,12 the eigenvalues
of the ideal SM system matrix Φe is estimated as follows:
̂ Ĉ Γ)
̂ − Γ(
𝜆(Φe ) = det{zI5 − [Φ ̂ −1 Ĉ Φ]},
̂

where |Ĉ Γ|̂ ≠ 0. The feedback matrix K = (Ĉ Γ) ̂ −1 Ĉ Φ


̂ is calculated using the MATLAB formula place, then the
first part of the problem is achieved. The second part is obtained by finding C from K. Assuming the design condition
̃ = 1, then the following relationship is defined:
|C Γ|
[ ]
c12 c14 c15 ⎡ Γ̂ 1 ⎤
Ĉ ΓF
̂ −1 = c22 c24 c25 , F = ⎢ Γ̂ 2 ⎥ .
c32 c34 c35 ⎢̂ ⎥
⎣ Γ3 ⎦
Assigning arbitrary values for c22 = c34 = c32 = 0, c14 = c15 = c25 = c35 = 1, c12 = 0.5, and c24 = 2, then
̂ −1 | = 1. Therefore, |F−1 | = 1, ie, F is a nonsingular matrix, then
|C ΓF
[ ]
0.5 1 −32.7756
Ĉ Γ̃ = 0 2 −66.1926
00 1
and
[ ]
−1.3493 −0.0003 −0.3478 0.0484 −0.1314
C = 10 4
−5.9083 −0.0015 −1.6902 0.1567 −0.4260 .
0.0385 0.0000 0.0094 −0.0014 0.0039
It is easy to check that the eigenvalues of the resultant equivalent ideal SM system matrix are exactly located on the
desired eigenvalues.

Solution 2 (Ackermann's approach).


In order to simplify the description of the Ackermann's procedure, let us consider an equivalent SM desired eigenval-
ues 𝜆d = {0.25, −0.8}, which are different from the previous solution. Considering the rearranged matrix W indicated
in (145) and following the approach explained in 3.1.2, the coefficients of the matrix C ∈ ℝ3×5 are obtained as follows.
• Find the reduced-order SM desired characteristic polynomial 𝛼(s). Therefore, we have

𝛼(s) = (s − 0.25)(s + 0.8) = s2 + 0.55s − 0.2.

• Compute Q = [qT1 qT2 qT3 ]T , where qi is defined in (46). Thus, we have


[ ]
−2066.458 −0.530 −633.997 37.948 −103.733
Q= −12.239 10.041 −4.647 0.186 0.244 .
−49.859 0.286 −39.734 −6.075 17.076

• Estimating fi using (48) and considering that C = [cT1 cT2 cT3 ]T , ci defined in (45), then the sliding surface coefficients
are as follows:
[ ]
−1.1985 −0.0065 −0.5080 −0.0187 0.0588
C = 1 × 10 −0.0122 0.0100 −0.0046 0.0002 0.0002 .
3

−0.0498 0.0003 −0.0397 −0.0061 0.0171


24 LUIS-DELGADO ET AL.

With this C, the equivalent matrix Φequ for the ideal SM mode is
⎡ −6.0466 −0.0016 −2.8455 −0.1238 0.3014 ⎤
⎢ −9.3793 −0.0024 −3.8613 −0.1648 0.4670 ⎥
Φequ = ⎢ 19.1611 0.0049 8.8716 0.3851 −0.9550 ⎥ ,
⎢ 375.1517 0.0961 154.0967 6.5731 −18.6803 ⎥
⎢ 160.5534 0.0412 67.2218 2.8758 −7.9957 ⎥
⎣ ⎦
where the eigenvalues are Λ(Φequ ) = {0, 0, 0, 0.25, − 0.8} = 𝜆d , ie, the SM is asymptotically stable with the spectrum
exactly equal to the designed values. Nevertheless, the definition of the matrix W given in (42) could result in sliding
surfaces with no real coefficients if the desired eigenvalues 𝜆d are complex. For example, if the poles must be set to
𝜆d = {0.7 ± 0.7i}, then it is not possible to find fi (Φ) with real valued coefficients. Therefore, matrix W(Φ) must be
differently defined. For example, choosing the following arrange:
[ ]
W = Γ1 ΦΓ1 Φ2 Γ1 Γ2 Γ3 , (146)
then the order of the subsystems are mi = {3, 1, 1} and, consequently, f1 (Φ) can be assigned with the pair of complex
eigenvalues 𝜆(Φequ ) = {0.7 ± 0.7i}. With the controllable matrix and desired poles, the matrix C is obtained as follows:
[ ]
−1.1985 −0.0066 −0.5080 −0.0187 0.0588
C = 1 × 10 −0.0122 0.0100 −0.0046 0.0002 0.0002 ,
3

−0.0499 0.0003 −0.0397 −0.0061 0.0171


and the equivalent matrix for the ideal SM mode is
⎡ −6.0466 −0.0016 −2.8455 −0.1238 0.3014 ⎤
⎢ −9.3793 −0.0024 −3.8613 −0.1648 0.4670 ⎥
̂ equ
Φ = ⎢ 19.1611 0.0049 8.8716 0.3851 −0.9550 ⎥ .
⎢ 375.1517 0.0961 154.0967 6.5731 −18.6803 ⎥
⎢ 160.5534 0.0412 67.2218 2.8758 −7.9957 ⎥
⎣ ⎦
This system matrix has the eigenvalues {0, 0, 0, 0.7 + 0.659i, 0.7 − 0.659i}, ie, the SM is asymptotically stable despite
̂ equ ), does not exactly correspond to 𝜆d .
the achieved spectrum, ie, 𝜆(Φ

Solution 3 (Regular representation approach).


Designing the sliding surface using the regular canonical representation implies obtaining the regular representation
of the system by a nonsingular transformation z(k) = Mx(k). Computing M using Equation (34) and applying the
nonsingular transformation, the following regular representation of the system is achieved:
⎡1 0 −0.27 0 0.10 ⎤ ⎡0 0 0⎤
⎢0 1 1062.76 0.70 −404.67 ⎥ ⎢0 0 0⎥
̂ = ⎢ 0 −0.212
Φ −113.03 −0.10 43.40 ⎥ , Γ̂ = ⎢ 1 0 0⎥. (147)
⎢ 0 −0.089 −10, 674.96 −6.05 4, 064.79 ⎥ ⎢0 1 0⎥
⎢ 0 −0.595 ⎥ ⎢0 1 ⎥⎦
⎣ −317.65 −0.27 121.93 ⎦ ⎣ 0
Thus, C parameters can be obtained estimating the feedback matrix of the reduced-order system presented in (36).
For this example, the following function in MATLAB was used:
C =[place(Phi_2_11,Phi_2_12,poles(4:5))eye(3)]* M,
̂ 11 and Phi_2_12 = Φ
where Phi_2_11 = Φ ̂ 12 . Then, the sliding surface is
[ ]
−22.30 −52.07 −1335514.03 67727.61 1133.62
C= 1.98 4.63 −0.138729.93 7035.36 118.06 . (148)
−58.56 −136.74 −3507590.36 177873.26 2977.33
From the equivalent SM dynamics, it can be shown that the poles are located on 𝜆d = {0, 0, 0, 0.7 ± 0.7i}.

Solution 4 (Cheng and Chang approach).


According to the methodology proposed in 3.1.3, the linear feedback matrix K is estimated such that the hypothetical
closed-loop system has the poles on some desired values 𝜆d . Choosing 𝜆d = {0.1, 0.1, 0.1, 0.7 ± 0.7i}, and by mean of
the MATLAB function place and applying (56), then the sliding surface is completely described as
[ ]
−23.8342 0.1045 70.7838 −4.2478 6.1844
s(k) = −2.7491 10.0443 8.1590 −0.4895 0.7124 x(k). (149)
−91.9335 0.2840 168.6704 −16.5259 14.7770
LUIS-DELGADO ET AL. 25

It can be proved that the spectrum of the equivalent SM is 𝜆(Φequ ) = {0, 0, 0, 0.7 ± 0.7i}. Using the mod-
ified approach proposed in the work of Draženović et al21 instead of (58) and using the MATLAB equation
C =[K eye(rank(Gamma))]* pinv([Fi Gamma]), then the sliding surface parameters are
[ ]
−21.4527 0.0786 63.4419 −3.1153 5.8129
C = −2.4745 9.0420 7.2959 −0.3635 0.6751 . (150)
−82.7468 0.2245 150.7845 −13.2766 14.2669

The eigenvalues of the ideal SM are Λ(Φequ ) = {0, 0, 0, 0.6633 ± 0.6479i}. These values do not match exactly the
desired ones.

Solution 5 (Lyapunov approach).


Following the Lyapunov-based approach proposed by Kim and Bandyopadhyay,66 the regular representation of
the systems is needed. As it was previously estimated, the next step consists in finding one matrix P that
solves the linear Lyapunov equation described in (71). Thus, by mean of the following MATLAB function,
P = dlyap((Phi_regular-Gamma_regular * K),Q), where Q = I5 and K is a feedback gain obtained to set
̂ in a stable location. In this example, K is obtained for the same poles 𝜆d used for the previous arbi-
the eigenvalues of Φ
trary eigenvalues solutions. Therefore, with matrix P obtained, the following sliding surface parameters are estimated
using (69):
[ ]
−71.916 0.018 −2, 751.099 138.994 6.140
C= −6.388 0.002 −257, 129.569 13039.776 218.221 . (151)
−188.843 −0.048 −7, 733.903 384.511 16.558

The poles of the SM dynamics matrix are located at {0, 0, 0, 0, 0.033}, ie, the SM is asymptotically stable.

Solution 6 (Optimal design of sliding surfaces approaches).


On the contrary to the previous eigenvalues assignment approaches, the following ones are based on optimal methods.
Considering the Lyapunov function V(k) = xT (k)Px(k), then, solving the Ricatti equation given in (79) for Q = In , the
optimal matrix P ∈ ℝn×n can be obtained using the dlqr MATLAB function. The P matrix can be used for obtaining
an optimal sliding surface via Spurgeon30 and Edwards et al32 sliding surface definition s(x) = ΓT Px(k), such that
[ ]
0.04492 −0.01145 0.51558 0.43938 0.17185
s(x) = 0.00144 0.09967 0.00110 0.00032 −0.00148 x(k).
−0.16845 −0.00002 −0.22193 −0.16844 −0.01196

Computing the equivalent control, then the equivalent system matrix for the SM is as follows:

⎡ 9.92541 0.00234 5.01517 0.22287 −0.44557 ⎤


⎢ 0.00239 1.00000 0.00175 0.00008 −0.00012 ⎥
Φequ = ⎢ 5.01517 0.00175 15.01468 0.67708 −0.25982 ⎥ , (152)
⎢ 0.22287 0.00008 0.67708 1.03276 −0.01159 ⎥
⎢ −0.44557 1.02225 ⎥⎦
⎣ −0.00012 −0.25982 −0.01159

where Λ(Φequ ) = {0, 0, 0, 0.65406 ± 0.64748i}.

Solution 7 (Min-max controller approach).


In absence of disturbances, the design method and the results are exactly the same as the previous solution.

Solution 8 (Luis et al approach).


Considering the proposed methodology proposed in the work of Luis-Delgado et al,39 then the first step consists on
finding the feedback matrix K. Using the following formula in MATLAB code:

K = place(Phi,Gamma,poles], (153)

where Phi = 𝚽, Gamma = 𝚪, and poles = 𝜆d (𝚽equ ), then the obtained feedback gain matrix is as follows:
[ ]
−23.834 −0.100 68.086 3.486 8.161
K= −2.749 10.019 7.677 0.397 1.009 .
−91.934 −0.3114 158.264 2.448 23.497
26 LUIS-DELGADO ET AL.

Choose an arbitrary matrix E1 defined as follows:


[ ]
111
E1 = 011 . (154)
001

Using (101), then the switching surface is completely defined, ie,


[ ]
−118.517 10.368 247.614 −21.2632 21.674
s(x) = −118.517 10.368 247.614 −21.2632 21.674 x(k).
−91.934 0.2833 168.670 −16.529 14.777

Using MATLAB function eig(), then it can be shown that the eigenvalues of the equivalent system matrix are
Λ(Φequ ) = {0.7 + 0.7i, 0.7 − 0.7i, 0, 0, 0}, therefore the poles are located exactly where it was specified.

Solution 9 (LMI approach).


Finally, for illustrating an LMI approach for the design of SM switching surface, we have chosen the proposition given
in the work of Nizar and Saïd,68 where the system dynamics is described as follows:

x(k + 1) = Φx(k) + Φd x(k − h) + Γu(k), (155)

where h is delay parameter and Φd ∈ ℝn×n . For this example, Φd = 0 because the time delay is not considered because
we have selected the nominal system obtained in the work of Xie et al.94 Following the LMI designing proposed in the
works of Nizar et al,68,70 the system must be transformed to the regular form, then the system representation given in
(161) is needed. Then, using theorem 1 in the work of Nizar et al,70 where it is proven that a system given in (155) is
asymptotically stable via DVSC if there exists symmetrical matrices P, Q ∈ ℝn×n such that the following LMI holds:
[ ]
W −L ∗ ∗
0 −L ∗ < 0, (156)
Φ11 L + Φ12 w1 Φd11 L + Φd12 w1 −L

where W = P−1 QP−1 , L = P−1 , w1 = −CL, and where (∗) denotes the transposed elements in the symmetric location.
The matrices Φ11 , Φ12 , Φd11 , Φd12 are given from the regular representation of (155).Therefore, the switching surface
is achieved as follows:
[ ]
C = −w1 L−1 Im M, (157)

where M is given in (34). Using the following equations of the MATLAB toolbox YALMIP,95 the feasible solution of
the LMI given (155) is achieved:
% variables statement

n = 5; m = 3; n1 = n − m;
L = sdpvar(n1); W = sdpvar(n1); w1 = sdpvar(m, n1);

% no-delay statement

Phi_d11 = zeros(n1, n1); Phi_d12 = zeros(n1, m);


% LMI definition

F = [W − L, zeros(n1, n1), L′ ∗ Phi_11′ + w1′ ∗ Phi_12′; zeros(n1, n1),


− L, L′ ∗ Phi_d11 + w1′ ∗ Phi_d12′; Phi_11 ∗ L
+ Phi_12 ∗ w1, Phi_d11 ∗ L + Phi_d12 ∗ w1, −L] < 0;

% Solving LMI

sol = solvesdp( F); W = double(W); L = double(L); w1 = double(w1).


LUIS-DELGADO ET AL. 27

Then, we have
[ ]
4.7347x106 0
L=
0 1.4831x106
[ ]
−0.0020 −0.1689
W = −0.1689 −0.0002 x10−4
[ ]
3.5215 0.000C_
w1 = −0.3128 −0.0001 x10−4 .
9.2477 0.0024

From (157), the sliding surface in the original state variable coordinate is as follows:
[ ]
74.376 −0.018 −2751.745 139.033 6.017
s(x) = 6.607 −0.002 −257, 128.804 13039.736 218.232 x(k).
195.315 −0.050 −7735.602 384.615 16.233

Considering the ideal SM equivalent dynamics, then Λ(Φequ ) ≈ {0, 0, 0, 0, 0}, thus the ideal SM is asymptotically
stable.
Another solution to the same problem where LMI approach are applied can be seen in the work of Tapia et al.22
This example presents a wide range of research approaches applied to the design of switching surface, which is very
useful to the process control designers because it gives a fruitful comparison among all of them. From this example, it
can be seen that the implementation of the different approaches is not very difficult. The criteria for choosing among
the wide range of options depend on every process to be controlled and the designer designing experience. It is worth
mentioning that, as far as we have known, there are no any research work including some numerical examples of
Ackerman's approach applied to MIMO systems presented in the work of Huang and Yeung61 ; hence, this example
is an important contribution of this review work. Some useful remarks that can easily seen from this example are
covered in 7. The next examples are more concerned to the process control research area.

Example 2. Continuous stirred tank reactors (CSTRs). They are very common processes in the industry, there-
fore they are widely used as design examples in nonlinear process control design.50,56,82,89,96 These chemical plants
exhibit a high nonlinear behavior and, for the cases of isothermal CSTR in which Van de Vusse reactions are acting,
nonminimun phase dynamics take place. To overcome this problem in the specification of a sliding surface, Sharaz
and Bozorgmehry Boozarjomehry56 proposed an I/O fuzzy approach where the surface parameters are obtained from
the output plant rather than the space state vector. On the other hand, the isothermal CSTR where Van de Vusse
reactions are present, more nonlinearities are acting compared with isothermal ones. Some approaches have been
proposed for the design of sliding switching surfaces considering these type of processes.56,96 Camacho and Smith50
proposed a scheme of continuous time slide mode control applied to chemical processes. In this approach, FOPDT
is used for modeling the plants. In this research work, a CSTR, where an exothermic reaction A → B takes place, is
controlled using a switching surface similar to the one given in (8), ie,

( )n−1 t
d
s(𝑦(t))
̃ = +𝜆 e(t)dt, (158)
dt ∫
0

where e(t) is the tracking error. Another examples are the pH CSTR, as the plant used in other works.56,97,98 In the VSC
approaches presented by Chen and Peng,89,96 a nonlinear normal canonical representation is used (see Section 4.1.3)
using a local transformation. Hence, a simple linear sliding surface was proposed to guarantee that the transformed
dynamics system to be asymptotically stable. In the work of Xie et al,94 LQR techniques are used for the design of an
optimal SM switching surface for a DVSC control of a discrete time MIMO linear system, which consists in a two-stage
chemical reactor with recycle, which is described by in the general delayed state space model, ie,

.
xc (t) = Ax(t) + Buc (t − 𝜏1 )
.
𝑦c (t) = Cxc (t − 𝜏2 ).
28 LUIS-DELGADO ET AL.

The work of Xie et al94 is applied for identifying the delay-free discrete-time model of the multiple time-delay process,
resulting in a delay-free controllable and observable discrete-time model of the form given in (40), where

⎡ 0.7319 −0.1356 0.2518 −0.0206 −0.0618 ⎤


⎢ −0.2266 0.6444 −0.2326 −0.0796 0.0348 ⎥
Φ = ⎢ −0.1877 0.2833 0.4511 −0.5658 −0.0514 ⎥ (159)
⎢ 0.0004 −0.0196 0.5607 −0.0928 0.5230 ⎥
⎢ −0.0263 −0.5448 ⎥⎦
⎣ −0.1163 −0.0275 −0.3962
⎡ −0.1507 −0.2441 ⎤
⎢ 0.0758 −0.2644 ⎥
Γ = ⎢ −0.2104 0.0476 ⎥ .
⎢ 0.1298 −0.0198 ⎥
⎢ 0.0305 −0.0231 ⎥⎦

The eigenvalues of the discretized system matrix are

Λ(Φ) = 𝜆d = {0.6110, 0.3992, 0.2553, −0.0378 ± 0.4204i}.

Supposing that a DVSC is implemented, if the switching surface is the one defined in (5), then the estimation of
the parameters of C ∈ ℝ2×5 can be obtained using the approaches here explained (for the eigenvalue placement
approaches, we have considered that the eigenvalues of the ideal SM system matrix Φequ must be shifted to the arbitrary
values Λd (Φequ ) = {0.2, 0.7 ± 0.7i}).
Solution 10 (Controller canonical approach).
As it was explained in the previous example, the nth-order linear independent columns controllable matrix W is
calculated such that
[ ]
W = Γ1 ΦΓ1 Φ2 Γ1 Γ2 ΦΓ2 . (160)
̂ Γ}
Then, using (160), the transformation matrix T can de obtained such that the following transformed system {Φ, ̂ is
achieved:
⎡ 0 1 0 0 0⎤
⎢ 0 0 1 0 0⎥
̂ = ⎢ 1.5333 −2.1088 1.7625 0.3399 −0.1721 ⎥
Φ
⎢ 0 0 0 0 1⎥
⎢ 2.3562 −4.8354 ⎥
⎣ 0 0.5151 −0.5727 ⎦
⎡ ⎤ ⎡
0 0 0 ⎤
⎢ 0⎥ ⎢0 0⎥
Γ̂ = ⎢ Γ̂ 1 ⎥ = ⎢ 1 0.0394 ⎥ .
⎢ 0⎥ ⎢0 0⎥
⎢ ̂ ⎥ ⎢0 1 ⎥⎦
⎣ Γ2 ⎦ ⎣
Following the procedure used in Example 1 and for the desired eigenvalues location 𝜆d = {0.2, 0.7 ± 0.7i}, then the
obtained matrix C ∈ ℝ2×5 is
[ ]
6.0980 −9.7471 −10.7246 4.6734 −23.1233
C = 12.2377 −18.3827 −16.2084 9.7777 −47.2710 .

From the ideal SM equivalent matrix Φequ , it can be verified that the eigenvalues are the desired eigenvalues 𝜆d .

Solution 11 (Ackermann's approach).


From the rearranged matrix W indicated in (160) and following the proposal explained in 3.1.2, the coefficients of the
matrix C ∈ ℝ2×5 are obtained as follows.
• The reduced-order SM desired characteristic polynomial 𝛼 (s) is

𝛼(s) = s3 − 1.6s2 + 1.26s − 0.196.

• Matrix Q defined in (46) is estimated such that


[ ]
−2.6223 3.0982 4.6548 2.8290 −0.5854
Q= 4.4007 −6.7119 −11.0820 −13.3902 18.9623 .
LUIS-DELGADO ET AL. 29

• Finally, for C = [cT1 cT2 ]T , where ci is defined in (45), then the sliding surface is equal to
[ ]
−1.5654 1.2256 −1.0256 4.7776 −5.4013
C= 5.4377 −8.6619 −8.1428 3.1216 −21.0622 .

From the equivalent system matrix, it is possible to verify that the ideal SM mode has the eigenvalues in the desired
values 𝜆d .

Solution 12 (Regular representation approach).


From (34) and applying the nonsingular transformation, then the following regular system representation is achieved:
⎡ 1.1145 1.4588 3.2085 0.2660 −0.8457 ⎤ ⎡0 0⎤
⎢ 0.1349 2.1867 1.6226 0.7376 −0.8401 ⎥ ⎢0 0⎥
̂ = ⎢ −0.2092
Φ 0.1577 1.2261 −0.3411 0.0827 ⎥ , Γ̂ = ⎢ 0 0⎥. (161)
⎢ 0.2213 0.7726 5.6366 −0.5604 −0.2578 ⎥ ⎢1 0⎥
⎢ 1.4308 ⎥ ⎢0 1 ⎥⎦
⎣ 6.0547 8.6327 2.1652 −2.7771 ⎦ ⎣
Thus, the C parameters can be obtained estimating the feedback matrix of the reduced-order system presented in (36).
Using the formula place in MATLAB, then the sliding surface is
[ ]
−1.5041 1.2744 −0.8945 4.8710 −4.7119
s(k) = −1.5569 −1.2388 −3.3272 −2.3689 −17.4843 x(k). (162)

From the equivalent SM dynamics, it can be shown that the poles are located on 𝜆d = {0, 0, 0.2, 0.7 ± 0.7i}.

Solution 13 (Cheng and Chang approach).


According to the methodology proposed in 3.1.3, the linear feedback matrix K is estimated such that the hypothetical
closed-loop system has the poles on some desired values 𝜆d . For 𝜆d = {0.1, 0.1, 0.2, 0.7 ± 0.7i} and using the MATLAB
function place and applying (56), then the sliding surface is completely described as follows:
[ ]
−1.2099 0.2520 −0.3966 4.2262 5.4611
s(k) = −1.5928 −1.9548 −0.7416 −0.7400 −4.9781 x(k). (163)

It can be proved that Λ(Φequ ) = {0, 0, 0.2, 0.7 ± 0.7i}. Using the modified approach proposed in the work of
Draženović et al21 instead of (58) and using the MATLAB equation C =[K eye(rank(Gamma))]* pinv([Fi
Gamma]), the sliding surface parameters are as follows:
[ ]
−1.0148 0.2842 −0.2495 4.1386 6.3247
C = −1.4076 −1.7712 −0.6527 −0.5690 −5.6206 . (164)
( )
The eigenvalues of the ideal SM are Λ Φequ = {0, 0, 0.3252, 0.7070 ± 0.7102i}. As it was achieved in the previous
example, the obtained eigenvalues do not match the desired ones.

Solution 14 (Lyapunov approach).


Following the Lyapunov-based approach proposed by Kim and Bandyopadhyay,66 the regular representation of the
systems is needed. As it was previously estimated, the next step consists in finding one matrix P that solves the linear
Lyapunov equation described in (71). Thus, by mean of the following MATLAB function,
P = dlyap((Phi_regular-Gamma_regular * K),Q),
̂ in a stable location. For the example,
where Q = I5 and where K is a feedback gain obtained to set the eigenvalues of Φ
K was obtained for the same poles 𝜆d used for the previous arbitrary eigenvalues solutions. Therefore, with matrix P
obtained, the sliding surface is estimated using (69) such that
[ ]
−2.2047 1.6896 −3.1491 −0.4409 −2.1527
C= 2.0724 −5.2636 −7.1223 −1.8219 −18.0576 . (165)

The poles of the SM dynamics matrix are located at {0, 0, 0.0676, 0.3228 ± 0.0781i}, ie, the SM is asymptotically
stable.

Solution 15 (Optimal design of sliding surfaces approaches).


For an optimal SM that minimizes the performance index given in (76) and the approach proposed in the works of
Spurgeon30 and Edwards et al,32 where s(x) = ΓT Px(k) can be used, such that P holds for Ricatti equation given in
30 LUIS-DELGADO ET AL.

(79), then, for Q = 12 I5 , the achieved sliding surface is as follows:


[ ]
−0.3840 −0.3148 −0.5550 0.2362 −0.0390
s(x) = −0.2306 −0.2924 0.0952 0.0802 −0.0607 x(k).

It can be shown that the eigenvalues of the equivalent system matrix are Λ(Φequ ) = {0, 0, 0.2805, − 0.1028 ± 0.1560i},
therefore the system is asymptotically stable.

Solution 16 (Luis et al approach).


The solution follows the methods used in the previous example. Using MATLAB, the feedback matrix K is
[ ]
−1.0101 −0.5038 1.6773 −2.3266 −0.6610
K = −0.4530 −0.6603 −0.5589 2.6490 2.3936 .

Using the following arbitrary design matrix definition:


[ ]
1.0 0.5
E = 0.0 0.1 ,

then using (101), the sliding surface achieved is as follows:


[ ]
−2.0063 −0.7254 −0.7674 3.8562 2.9720
s(x) = −0.1593 −0.1955 −0.0742 −0.0740 −0.4978 x(k).

Solution 17 (Min-max controller approach).


In absence of disturbances, the design method and the results are exactly the same as the previous solution.

Solution 18 (LMI approach).


Following the same methodology applied to the previous example, then LMI approach used for solving this problem
gives the following solution:
[ ]
1.5222 0.3416 0.2153
L = 0.3416 1.4831 −0.2279
0.2153 −0.2279 1.7011
[ ]
−0.3226 0.3416 0.2153
W = 0.3416 −0.3617 −0.2279
0.2153 −0.2279 −0.1436
[ ]
2.1878 −1.1796 7.1642
w1 = 3.8123 2.0771 8.7891 .

Then, from (157), the sliding surface is as follows:


[ ]
−0.9549 0.3947 −4.0377 0.5831 −3.2474
s(x) = −1.3353 −1.9002 −5.2523 −5.8420 −13.2454 x(k).

Considering the ideal SM equivalent dynamics, then Λ(Φequ ) = {0, 0, 0, 0, 0.8216}, thus the ideal SM is asymptotically
stable.

Example 3. Interconnected tanks are processes that can be frequently found in the industry. For this reason, there is
a lot of research works where DVSC-SM and VSC-SM techniques are applied for controlling theses processes. Some
examples can be seen in other works.86,99-102 In the work of Iqbal et al,99 three tanks nonlinear systems are considered
for VSC controlling, where it is proposed a linear sliding surface s (t) = Cx(t). The two interconnected tanks considered
in the work of Xi and Hesketh86 is a second-order nonlinear SISO that we have selected for illustrating the nonlinear
methodologies explained in this work. As it can be seen in Figure 3, the system consists in a double tank system
connected in cascade and where water is pumped into the tank T1 . Defining to h1 (t) as the difference between the two
tanks levels and h2 (t) is the level of the tank T2 , then the continuous time dynamics are given as follows:
. √
h1 (t) = −p1 h1 (t) + qu(t) (166)
. √ √
h2 (t) = p1 h1 (t) − p2 h2 (t),
LUIS-DELGADO ET AL. 31

FIGURE 3 Scheme of the two tanks connected process

where p1 , p2 denote the output flow coefficients including gravity and density for T1 and T2 , respectively; q corresponds
to the input flow coefficient; and u(t) is the input water flow rate.

Solution 19 (Discrete-time canonical controller form approach).


Let us obtain a discrete-time representation using Euler's method. Therefore, for the sampling time T, the following
discrete-time representation is obtained:
[ ] [ √ ] [ ]
h1 (k + 1) h1 (k)√
− p1 T h1 (k)√ qT
h2 (k + 1) = h2 (k) + p1 T h1 (k) − p2 T h2 (k) + 0 u(k).

For the sake of simplicity, let us consider that p1 = p2 = p. Then, defining the following state variable x1 (k) = h1 (k),
x2 (k) = h2 (k) − h2r , and the output y(k) = h2 (k), the local generalized controller canonical form (LGCCF) can be
found such that
[ ] [ √ ]
qT x1 (k)(− pT x1 (k) )
g= 0 , 𝑓 = √ √
x2 (k) + h2r + pT x1 (k) − x2 (k) + h2r
(√ √ )
L𝑓 h = x2 (k) + h2r + pT x1 (k) − x2 (k) + h2r , Lg h = 0
pT qpT 2
L2𝑓 h = 1 − √ , Lg L𝑓 h = √ .
2 x2 (k) + h2r 2 x1 (k)
It can be seen that the relative order is r = 2. Since the relative degree is equal to the order of the system, then there
is no zero dynamics. Hence, the system is stabilizable. The following discrete-time coordinate transformation can be
used:
[ ] [ x2 (k)
]
𝜉1 (k) (√ √ )
𝜉2 (k) = x2 (k) + h2r + pT , (167)
x1 (k) − x2 (k) + h2r
and the corresponding inverse transformation is as follows:
[ ] [ (√ )2 ]
1
x1 (k) 𝜉 1 (k) + h 2r + (𝜉 2 (k) − 𝜉 1 (k) − h )
2r pT .
x2 (k) = 𝜉1 (k)
Consequently, a continuous time LGCCF dynamics are obtained as follows:
[ ] [ ]
𝜉1 (k + 1) 𝜉2 (k)
𝜉 (k + 1)
2
= a (𝜉(k)) + b (𝜉(k)) u(k) , (168)

where
pT
a(𝜉(k)) = 1 − √
2 x2 (k) + h2r
qp2 T 3
b(𝜉(k)) = ( √ ).
2 pT 𝜉1 (k) + h2r + (𝜉2 (k) − 𝜉1 (k) − h2r )

Then, a linear switching surface can be defined as follows:


s(𝜉) = c1 𝜉1 (k) + 𝜉2 (k).
Thus, the ideal SM dynamics are defined as follows:
𝜉1 (k + 1) = −c1 𝜉1 (k).
32 LUIS-DELGADO ET AL.

FIGURE 4 State space trajectories (𝜉 1 , 𝜉 2 ) for two surface specifications c = 0.1 (blue ones) and c = 0.4 (red ones) for the controller
canonical form approach [Colour figure can be viewed at wileyonlinelibrary.com]

FIGURE 5 Comparative between the values of s(k) and 𝜉 1 (k) for two c1 parameters values (c1 = 0.1 and c1 = 0.9) for the controller
canonical form approach [Colour figure can be viewed at wileyonlinelibrary.com]

Hence, if |c1 | < 1, then an asymptotically stable SM is achieved.


The parameter values are the ones used in the work of Xi and Hesketh,86 T = 0.1 s, p = 1.15 × 10−2 , q =
0.99 × 10−3 m∕(Vs), and h2r = 0.1 m for numerical simulations. Let us consider a DVSC where the control law is
obtained via the reaching law approach2 such that

u(k) = uequ (k) − g(k)−1 [s(k)(𝜇T − 1) + 𝜀Tsign(s(k))], (169)

such that the arbitrary designing parameters 𝜀, 𝜇 must satisfy the following restriction:
𝜀T
Δ= , (170)
1 − 𝜇T
where 2Δ corresponds to the width of quasi-SM band (QSMB). For a QSMB < 0.2 and choosing 𝜇 = 5, then 𝜀 = 0.5.
Using the tool Simulink of MATLAB, the controlled system dynamics are simulated. For the simulations, three values
for c1 parameter are selected. In Figure 4, it can be seen the state variable trajectories for c1 = 0.1 and c1 = 0.4,
and in Figure 5, the output of the system y(k) = 𝜉 1 (k) and the switching surface s(k) values are shown for c1 = 0.1
and c1 = 0.9. We have chosen c1 = 0.9 in Figure 5 for the purpose of illustrating the change of the amplitude of
the oscillations of the state states in function of the switching surface parameters. However, the representation of the
state trajectories for values of c1 greater than 0.4 made the picture a little difficult to appreciate the lines, therefore we
preferred to use c1 = 1 and c1 = 0.4 in Figure 4.

Solution 20 (Discrete-time regular canonical form approach).


Let us start finding the continuous time controller canonical form. Considering the state variables x1 (t) = h1 (t),
x2 (t) = h2 (t) − h2r , and y(t) = h2 (t), then it is not difficult to verify that the relative order is r = 2 and the following
coordinate transformation can be applied:
[ ] [ x2 (t)
]
𝜉1 (t) (√ √ )
𝜉2 (t) = p , (171)
x1 (t) − x2 (t) + h2r
LUIS-DELGADO ET AL. 33

and the corresponding inverse transformation:


[ ] [ (√ )2 ]
x1 (t) 𝜉 1 (t) + h 2r + 𝜉 2 (t)∕p .
x2 (t) = 𝜉 (t)1

Then, the continuous time LGCCF dynamics are as follows:


[. ] [ ]
𝜉. 1 (t) 𝜉 (t)
= a(𝜉(t)) +2b(𝜉(t))u(t) , (172)
𝜉 2 (t)
where
p2 p 𝜉2 (t)
a(𝜉(t)) = − − √
2 2 𝜉1 (t) + h2r
qp
b(𝜉(t)) = (√ ).
2 𝜉1 (t) + h2r + 𝜉2 (t)∕p

Using again the Euler's method for the discretization of the system given in (172) with a sampling time T, then
discrete-time system dynamics are as follows103 :
[ ] [ ] [ ]
𝜉1 (k + 1) 𝜉1 (k) + T𝜉2 (k) 0
𝜉 (k + 1)
2
= 𝜉 (k) + Ta (𝜉(k))
2
+
2 Tb(𝜉(t)) u(k). (173)

For this canonical representation, the state variable 𝜉 1 (k) corresponds to the discrete error x2 (k) = h2 (k) − h2r and
𝜉 2 (k) is the discrete-time approximation to the time derivative of this error. Choosing the sampling time as T = 0.1 s,
then the SISO nonlinear plant is completely defined, and considering only the nominal plant, ie, f (k) = 0, then

𝜉1 (k + 1) = 𝜉1 (k) + 0.1𝜉2 (k) (174)


𝜉2 (k + 1) = m(k) + n(k)u(k),

where
5.75 × 10−4 𝜉 (k)
2 2
m(k) = 𝜉2 (k) − √ − 6.6125 × 10−6
0.1 + 𝜉1 (k)
( )
5.692 × 10−7
n(k) = √ .
86.9565𝜉2 (k) + 0.1 + 𝜉1 (k)
Defining the linear sliding surface
s(k) = c1 𝜉1 (k) + 𝜉2 (k),
then, for the SM
𝜉2 (k) = −c1 𝜉1 (k),
hence the equivalent dynamics is given by

𝜉1 (k + 1) = 𝜉1 (k) (1 − 0.1c1 ) ,

therefore the sliding surface is asymptotically stable if 0 < c1 < 10. As in the previous methodology applied to
this example and using the tool Simulink of MATLAB, the controlled system dynamic was obtained. In Figure 6, it
is shown two phase diagrams corresponding to two surface definitions: one for c1 = 2 and the other for c1 = 5,
where it is clear that both SM dynamics are asymptotically stable around 𝜉(k) = 0. In Figure 7, the system output
y(k) = 𝜉 1 (k), and the surface value s(k), it is shown for the same two surface definitions (c1 = 2 and c1 = 5).

Solution 21 (Lyapunov approach).


Let us consider the following control law, which stabilizes the two tank systems described in (166), as follows:

p x2 (t) x1 (t)
u(t) = k(x(t)) = − , (175)
q x1 (t)
34 LUIS-DELGADO ET AL.

FIGURE 6 State space trajectories (𝜉 1 , 𝜉 2 ) for two surface definitions: c = 2 (blue ones) and c = 5 (red ones) for the regular canonical
form approach solution [Colour figure can be viewed at wileyonlinelibrary.com]

FIGURE 7 Comparative between the values of s(k) and 𝜉 1 (k) for two c1 parameters values (c1 = 2 and c1 = 5) for the regular canonical
form approach solution [Colour figure can be viewed at wileyonlinelibrary.com]

then the closed-loop system dynamics are


[. ] ⎡ x (t)−x (t)
−p 1√ 2 ⎤
x. 1 (t) ⎢ ( x1 (t) )⎥.
= √ √
x2 (t) ⎢p x1 (t) − x2 (t) + h2r ⎥⎦

Considering the following Lyapunov function:

V(x(t)) = x12 (t) + x22 (t), (176)

then
. ( √ √ )
V(𝜉(t)) = −2p x1 x1 (t) + x2 x2 (t) + h2r < 0,

where the conditions x1 (t) ≥ 0 and x2 (t) ≥ 0 are considered. Hence, the system is asymptotically stable, therefore
(176) is an admissible Lyapunov function. Using (140),
[ ][ ]T
s(x) = q 0 2x1 2x2 = 2qx1 (t). (177)

Therefore, from (177), the SM dynamics are x1 (t) = 0.

7 OV ERALL REMARKS AND PROCESS CO NTRO L GUIDANCE

Before the final conclusions, it was considered that it is important to highlight some details that can be made about the
methods previously explained.
• The eigenvalues assignment approaches provide easier design methodologies regarding to linear system and where the
canonical representation ones are the most widely used. Nevertheless, finding similar transformation for high-order
LUIS-DELGADO ET AL. 35

multi-input systems tends to be complicated. Ackermann's approach18,61 gives an explicit formula to the sliding sur-
face, setting the SM eigenvalues on the desired values. Using the algorithm proposed by Dorling and Zinober62 provides
a robust eigenvalue assignment with the drawback that an iterative computing method must be implemented. The
proposal presented by Chen and Chang64 and the modified version introduced by Draženović et al21 give an easy
explicit definition of the sliding surface. It is important say that the eigenvalues must be carefully chosen because the
performance of the steady state systems dynamics are determined by them.
• The linear Lyapunov approach presented by Su et al28 proposes a sliding surface explicit equation, which directly gives
the switching surface coefficients. Nevertheless, applying this method, a feedback control matrix K that makes the
system asymptotically stable via eigenvalue assignment is needed. The ideal SM achieved by this method is asymptoti-
cally stable but the SM eigenvalues cannot be arbitrary set. In contrast with this method, the Lyapunov-based method
presented by Kim and Bandyopadhyay66 can be directly applied to discrete-time systems. However, one disadvantage
of the methodology presented by Kim and Bandyopadhyay66 compared with the work of Su et al28 is that the system
dynamics must be given in regular form.
• The method of switching surface design applied to ISMC34,35 is not an involved methodology and it can be used in the
cases of perturbed linear systems where the disturbances bounds are known.
• The design methodology proposed in the work of Luis-Delgado et al39 unifies, in a single scheme, the design of switching
surfaces applied to discrete-time linear MIMO system via either pole placement or LQR, with high accuracy and with-
out the use of any coordinate transformation. These features make the designing procedure an easier task compared
with other methodologies reducing the design efforts.
• Optimal approaches applied to linear systems have the advantage that the SM eigenvalues are not explicitly needed
to be previously estimated.13,24,30,32,62,73 The main advantage of these approaches is that they improve the performance
of the state dynamic on the sliding surface. The weakness of these approaches is the difficulty in finding the right
weighting factors, ie, the coefficients of the matrices Q and R in (72).
• The min-max developments presented by Lai et al31 are oriented to output feedback SM approaches, which is very
useful when the system state is not attainable. This methodology is not complicated and yields to an optimal SM. The
sliding surface definition presented in the work of Lai et al31 is identical to another one presented by Spurgeon,30 which
is achieved in a different way.
• The LMI has been becoming an extended useful tool that is currently applied in almost every design method where
Lyapunov or optimal schemes are requested.22-26,67-69,75,77-79 The methods presented by Nizar and Saïd,68,79 Yan et al,69
and Pai78 are directly based on LMI and give an optimal solution to the VSC applied to disturbed linear systems with
delays.
• The methodologies proposed by Luis-Delgado et al39,40 are the only ones that allows the designers to adjust both
reaching and nonideal SM by the use of a design matrix.
• As far as nonlinear system is concerned, the Lyapunov-based sliding surface design method is a good approach if
the stability of the controlled system is the main objective of the DVSC. However, the second Lyapunov methodol-
ogy has the difficulty of finding feasible Lyapunov functions, which is not an easy task.44 The scheme proposed in
the work of Luis-Delgado et al40 integrates the linearizing procedure with the surface scheme in a unique equation,
which leads to a powerful design approach. Canonical-based methods are also used to design switching surface applied
to nonlinear systems but, in contrast to the linear case, there is no single general solution because every system
yields to different canonical representations, ie, the canonical forms have the same canonical structure but the non-
linear functions in those schemes depend of every case. Consequently, the resulting surfaces depend on every single
case too.

As design tips, it is important to be concern about the following issues.

• It is important to evaluate the need of linearizing the plant to use linear surface design approach. This task could become
very complex and it could not be directly implemented if the system dynamics have nonminimal phase behavior. These
drawbacks can be overcome including additional control implementation.96
• The decision of designing linear or nonlinear surface must be considered to simplify the design procedure or for
obtaining a better performance of the controlled process.28,45
• For multi-input nonlinear systems, the design of switching surface could imply a big endeavor. Therefore, considering
the robustness and the invariance property of SM controllers, using more simple dynamics models, is very helpful to
obtain a suitable solution.41,50
36 LUIS-DELGADO ET AL.

TABLE 1 Main features for the research proposals explained in this work
Approaches Weakness and Strengths
Canonical pole assignment13,16 Need canonical representation
Complex for high order systems
Ackermann's method18 Easy implementation
Only for SISO linear systems
Extended Ackermann's method61 Easy implementation
A priory poles location is needed
Robust pole assignment
Kautsky's algorithm42 Complex algorithm
Only for linear systems
Easy implementation
Chen and Chan algorithm64 Explicit surface definition
Only for linear systems
Robust SM definition
ISMC35 Explicit surface definition
Only for linear systems
Lyapunov28,66 Require Regular representation,
poles location and weighting factors
Need weighting factors
Optimal approach13,24,32,73 Explicit surface definition
Only for linear systems
Useful for perturbed delayed systems
LMI22,25,32,68 Need weighting factors
Only for linear systems
Simple and accurated
Luis et al approach39,40 Optimal and eigenvalue criteria
Sliding band adjustment
Only for some class of nonlinear systems
Useful for perturbed delayed systems
Min-max control31 Need weighting factors
Only for linear systems
Abbreviations: ISMC, integral sliding mode control; LMI, linear matrix inequality;
SISO, single-input–single-output; SM, sliding mode.

• Despite of the fact that DVSC were thought to be a state vector control scheme, some input-output schemes can be
used for designing sliding surfaces.26,104
• Choosing among the designing approaches here included is not an easy task. For process control, the nonlinear
approaches are the most useful tools. Using system dynamic transformation is the most easy way to design the switch-
ing surface considering that, in some cases, linear approaches can be used. Indeed, some local stable SM behavior can
be easily achieved with simple dynamics models of the plant.
On Table 1, the main researching highlights related to the design of sliding hyper-surfaces here explained can be seen.
It is not very difficult to note the high lack between linear and nonlinear research works.

8 CO N C LUSION S

This review article has traveled through the research works that have achieved relevant progresses on the problem of
designing sliding switching surfaces, starting from the very beginning, where the first milestones were established, until
the current works.
On the whole, the research works that can be found in the literature cover a large number of methods for the design of
sliding surfaces. We have considered a widespread range of linear methods, such as arbitrary eigenvalue assignment, opti-
mal sliding surface designing, LMI-based algorithm, and Lyapunov equation approach, as well as nonlinear approaches
such as canonical dynamics form and Lyapunov equation approaches. Currently, some of these algorithm are seldom
used, but some of them, such that LMI-based schemes, are becoming widely applied in several DVSC-SM research works.
LUIS-DELGADO ET AL. 37

On the other hand, from this review, it is evident that the number of research works related to the design of asymptoti-
cally stable sliding manifolds applied to nonlinear systems are not as extensive as the linear ones. Hence, further relevant
research works related with this area could be very useful because of the very large number of nonlinear system that can
be found in real life.

ACKNOWLEDGEMENTS
This work has been funded by the Spanish Ministry of Science, Innovation and Universities through the (COGDRIVE
project: Artificial Intelligence Techniques and Assistance to Autonomous Navigation) (DPI2017-86915-C3-3-R).

ORCID

José Darío Luis-Delgado https://orcid.org/0000-0002-2008-4067

REFERENCES
1. Furuta K. Sliding mode control of a discrete system. Syst Control Lett. 1990;14(2):145-152.
2. Gao W, Wang Y, Homaifa A. Discrete-time variable structure control systems. IEEE Trans Ind Electron. 1995;42(2):117-122.
3. Sira-Ramírez H. Non-linear discrete variable structure systems in quasi-sliding mode. Int J Control. 1991;54(5):1171-1187.
4. Milosavljević Č. General conditions for the existence of a quasi-sliding mode on the switching hyperplane in discrete variable structure
systems. Autom Remote Control. 1985;46(3):307-314.
5. Potts RB, Yu X. Discrete variable structure system with pseudo-sliding mode. J Aust Math Soc Ser B Appl Math. 1991;32(4):365-376.
6. Yu X, Potts RB. A class of discrete variable structure systems. In: Proceedings of the 30th IEEE Conference on Decision and Control;
1991; Brighton, UK.
7. Yan Y, Yu X, Yu S. Quantization effect on sliding-mode control of a second-order dynamical system. Paper presented at: 2012 12th
International Workshop on Variable Structure Systems; 2012; Mumbai, India.
8. Yan Y, Yu X, Yu S. Quantization effect on the discrete-time sliding-mode control systems. In: Proceedings of the 32nd Chinese Control
Conference; 2013; Xi'an, China.
9. Bartoszewicz A, Latosiński P. Discrete time sliding mode control with reduced switching-a new reaching law approach. Int J Robust
Nonlinear Control. 2016;26(1):47-68.
10. Jedda O, Douik A. Discrete-time sliding mode control for an inverted pendulum system. Paper presented at: 2018 International
Conference on Advanced Systems and Electric Technologies (IC_ASET); 2018; Hammamet, Tunisia.
11. Repecho V, Biel D, Arias A. Fixed switching period discrete-time sliding mode current control of a pmsm. IEEE Trans Ind Electron.
2018;65(3):2039-2048.
12. Utkin V. Variable structure systems with sliding modes. IEEE Trans Autom Control. 1977;22(2):212-222.
13. Utkin V, Yang KD. Methods for construction of discontinuity planes in multidimensional variable structure systems. Autom Remote
Control. 1978;39(6):1466-1470.
14. Dorling C, Zinober ASI. Two approaches to hyperplane design in multivariable variable structure control systems. Int J Control.
1986;44(1):65-82.
15. Dorling C, Zinober ASI. Computer Aided Design of Robust Multivariable Structure Control Systems. In: Computer Aided Design in Control
and Engineering Systems. Oxford, UK: Pergamon Press; 1986:353-358.
16. Hung JY, Gao W, Hung JC. Variable structure control: a survey. IEEE Trans Ind Electron. 1993;40(1):2-22.
17. Gao W, Hung JC. Variable structure control of nonlinear systems: a new approach. IEEE Trans Ind Electron. 1993;40(1):45-55.
18. Ackermann J, Utkin V. Sliding mode control design based on Ackermann's formula. In: Proceedings of 1994 33rd IEEE Conference on
Decision and Control - Volume 4; 1994; Lake Buena Vista, FL.
19. Perruquetti W, Barbot J-P. Sliding Mode Control in Engineering. New York, NY: Marcel Dekker, Inc; 2002.
20. Utkin V, Guldner J, Shi J. Sliding Mode Control in Electro-Mechanical Systems. Boca Raton, FL: Taylor and Francis; 1999.
21. Draženović B, Milosavljević Č, Veselić B. Comprehensive approach to sliding mode design and analysis in linear systems.
In: Bandyopadhyay B, Janardhanan S, Spurgeon SK, eds. Advances in Sliding Mode Control: Concept, Theory and Implementation. Berlin,
Germany: Springer-Verlag Berlin Heidelberg; 2013:1-19.
22. Tapia A, Márquez R, Bernal M, Cortez J. Sliding subspace design based on linear matrix inequalities. Kybernetika. 2014;50(3):436-449.
23. Choi HH. A new method for variable structure control system design: a linear matrix inequality approach. Automatica.
1997;33(11):2089-2092.
24. Kim KS, Park Y, Oh SH. Designing robust sliding hyperplanes for parametric uncertain systems: a Riccati approach. Automatica.
2000;36(7):1041-1048.
25. Choi HH. An LMI-based switching surface design method for a class of mismatched uncertain systems. IEEE Trans Autom Control.
2003;48(9):1634-1638.
38 LUIS-DELGADO ET AL.

26. Edwards C, Spurgeon S. Linear matrix inequality methods for designing sliding mode output feedback controllers. IEE Proc Control
Theory Appl. 2003;150(5):539-545.
27. Lin Y, Shi Y, Burton R. Modeling and robust discrete-time sliding-mode control design for a fluid power electrohydraulic actuator (EHA)
system. IEEE/ASME Trans Mechatron. 2013;18(1):1-10.
28. Su WC, Drakunov SV, Özgüner Ü. Constructing discontinuity planes for variable structure systems - a Lyapunov approach. In:
Proceedings of 1994 American Control Conference - Volume 1 (ACC); 1994; Baltimore, MD.
29. Castaños F, Fridman L. Dynamic switching surfaces for output sliding mode control: an ℋ ∞ approach. Automatica.
2011;47(9):1957-1961.
30. Spurgeon SK. Hyperplane design techniques for discrete-time variable structure control systems. Int J Control. 1992;55(2):445-456.
31. Lai NO, Edwards C, Spurgeon S. On discrete time output feedback sliding mode like control for non-minimum phase systems. Paper
presented at: 42nd IEEE International Conference on Decision and Control; 2003; Maui, HI.
32. Edwards C, Herrmann G, Hredzak B, Venkataramanan V, Spurgeon S. A discrete-time sliding mode scheme with constrained inputs.
Paper presented at: 2007 46th IEEE Conference on Decision and Control; 2007; New Orleans, LA.
33. Utkin V, Shi J. Integral sliding mode in systems operating under uncertainty conditions. In: Proceedings of 35th IEEE Conference on
Decision and Control; 1996; Kobe, Japan.
34. Abidi K, Xu J-X, Yu XH. On the discrete-time integral sliding mode control. Paper presented at: International Workshop on Variable
Structure Systems; 2006; Sardinia, Italy.
35. Abidi K, Xu J-X, Xinghuo Y. On the discrete-time integral sliding-mode control. IEEE Trans Autom Control. 2007;52(4):709-715.
36. Bai J, Lu R, Wu Z, Zhang R, Zhao X, Xue A. Robust ℋ∞ control of discrete-time singular systems via integral sliding surface. Asian J
Control. 2018;20(3):1296-1302.
37. Bandyopadhyay B, Deepak F. A nonlinear sliding surface for discrete-time uncertain systems. Paper presented at: 2010 11th International
Workshop on Variable Structure Systems (VSS); 2010; Mexico City, Mexico.
38. Patil M, Bandyopadhyay B. Discrete-time sliding mode control for state delay systems using nonlinear sliding surface. Paper presented
at: 2011 11th International Conference on Control, Automation and Systems (ICCAS); 2011; Gyeonggi-do, South Korea.
39. Luis-Delgado JD, Al-Hadithi BM, Jiménez A. Design of switching hyperplanes for multi-inputs multi-outputs discrete-time linear
systems. IET Control Theory Appl. 2016;10(2):126-135.
40. Luis-Delgado JD, Al-Hadithi BM, Jiménez A. A novel method for the design of switching surfaces for discretized MIMO nonlinear
systems. Int J Appl Math Comput Sci. 2017;27(1):5-17.
41. Sira-Ramírez H. Variable structure control of nonlinear systems through simplified uncertain models. Paper presented at: 1986 25th IEEE
Conference on Decision and Control; 1986; Athens, Greece.
42. DeCarlo R, Zak SH, Matthews GP. Variable structure control of nonlinear multivariable systems: a tutorial. Proc IEEE. 1988;76(3):212-232.
43. Spurgeon S, Davies R. A nonlinear design approach for sliding mode control systems. In: Proceedings of 32nd IEEE Conference on
Decision and Control - Volume 2; 1993; San Antonio, TX.
44. Su WC, Drakunov SV, Özgüner Ü. Constructing discontinuity surfaces for variable structure systems: a Lyapunov approach. Automatica.
1996;32(6):925-928.
45. Ghaffari A, Yazdanpanah MJ. Nonlinear sliding surfaces; computing and existence of solution. Paper presented at: 2008 International
Conference on Control, Automation and Systems; 2008; Seoul, South Korea.
46. Zhang X, Wang P, Yan M, Ju Y. Discrete-time sliding mode control of nonlinear time-delay systems based on TS fuzzy model. Paper
presented at: 2010 International Conference on Intelligent Control and Information Processing; 2010; Dalian, China.
47. Dong R, Shi DW. Optimal sliding mode design for nonlinear discrete-time systems. Paper presented at: 2011 IEEE International
Conference on Computer Science and Automation Engineering; 2011; Shanghai, China.
48. Nadzinski G, Vladev G, Zheng Y. A design of discrete-time SMC for nonlinear systems based on fuzzy TS model. Paper presented at: 2012
6th IEEE International Conference Intelligent Systems; 2012; Sofia, Bulgaria.
49. Slotine JJE, Li WL. Applied Nonlinear Control. Englewood Cliffs, NJ: Prentice Hall; 1991.
50. Camacho O, Smith CA. Sliding mode control: an approach to regulate nonlinear chemical processes. ISA Trans. 2000;39(2):205-218.
51. Moradi H, Saffar-Avval M, Bakhtiari-Nejad F. Sliding mode control of drum water level in an industrial boiler unit with time varying
parameters: a comparison with H∞ -robust control approach. J Process Control. 2012;22(10):1844-1855.
52. Li Y, Zhi-ming D, Bing-xiang Z, Ai-hua L. Application of the fuzzy discrete sliding variable-structure system on switched reluctance
motor. Paper presented at: 9th IEEE International Conference on Cognitive Informatics (ICCI); 2010; Beijing, China.
53. Pati A, Singh S, Negi R. Sliding mode controller design using PID sliding surface for half car suspension system. Paper presented at: 2014
Students Conference on Engineering and Systems; 2014; Allahabad, India.
54. Shihong Q, Samuel A. Analysis of robustness sliding mode control method for active suspension system. Int J Autom Control Eng.
2013;2(3):137-142.
55. Åström KJ, Wittenmark B. Computer-Controlled Systems: Theory and Design. 3rd ed. Upper Saddle River, NJ: Prentice-Hall International
Editions; 1997.
56. Shahraz A, Bozorgmehry Boozarjomehry R. A fuzzy sliding mode control approach for nonlinear chemical processes. Control Eng Pract.
2009;17(5):541-550.
57. Luenberger DG. Canonical forms for linear multivariable systems. IEEE Trans Autom Control. 1967;12(3):290-293.
LUIS-DELGADO ET AL. 39

58. Richter S, Lefebvre S, DeCarlo R. Control of a class of nonlinear systems by decentralized control. IEEE Trans Autom Control.
1982;27(2):492-494.
59. Chern TL, Chuang CW, Jiang RL. Design of discrete integral variable structure control systems and application to a brushless DC motor
control. Automatica. 1996;32(5):773-779.
60. Ackermann J, Utkin V. Sliding mode control design based on Ackermann's formula. IEEE Trans Autom Control. 1998;43(2):234-237.
61. Huang YJ, Yeung KS. Arbitrary eigenvalue assignment via switching hyperplanes design scheme and extension of Ackermann's formula.
Paper presented at: 1993 IEEE Region 10 International Conference on Computers, Communications, Control and Power Engineering
(IEEE TENCON); 1993; Beijing, China.
62. Dorling C, Zinober A. Robust hyperplane design in multivariable variable structure control systems. Int J Control. 1988;48(5):2043-2054.
63. Kautsky J, Nichols NK, Van Dooren P. Robust pole assignment in linear state feedback. Int J Control. 1985;41(5):1129-1155.
64. Chen YP, Chang JL. A new method for constructing sliding surfaces of linear time-invariant systems. Int J Syst Sci. 2000;31(4):417-420.
65. Castaños F, Fridman L. Analysis and design of integral sliding manifolds for systems with unmatched perturbations. IEEE Trans Autom
Control. 2006;51(5):853-858.
66. Kim KS, Bandyopadhyay B. A Lyapunov approach for discrete-time sliding hyperplane design and robust sliding mode control using
MROF. IFAC Proc Vol. 2008;41(2):1129-1134.
67. Choi HH. LMI-based sliding surface design for integral sliding mode control of mismatched uncertain systems. IEEE Trans Autom
Control. 2007;52(4):736-742.
68. Nizar A, Saïd NA. A new sliding surface for discrete second order sliding mode control of time delay systems. Paper presented at:
International Multi-Conference on Systems, Sygnals and Devices; 2012; Chemnitz, Germany.
69. Yan M, Mehr AS, Shi Y. Discrete-time sliding-mode control of uncertain systems with time-varying delays via descriptor approach.
J Control Sci Eng. 2008;2008. Article No 10.
70. Nizar A, Ltaief M, Saïd NA. A new discrete sliding mode control for systems with time varying delay: multi-structure approach. Paper
presented at: 10th International Multi-Conferences on Systems, Signals and Devices (SSD); 2013; Hammamet, Tunisia.
71. Yan M, Shi Y. Robust discrete-time sliding mode control for uncertain systems with time-varying state delay. IET Control Theory Appl.
2008;2(8):662-674.
72. Jung S-L, Huang H-S, Tzou Y-Y. Self-tuning discrete sliding mode control of a closed-loop regulated PWM inverter with optimal sliding
surface - Volume 2. Paper presented at: 27th Annual IEEE Power Electronics Specialists Conference (PESC); 1996; Maggiore, Italy.
73. Pieper JK. First order dynamic sliding mode control - Volume 3. In: Proceedings of the 37th IEEE Conference on Decision and Control;
1998; Tampa, FL.
74. VanAntwerp JG, Braatz RD. A tutorial on linear and bilinear matrix inequalities. J Process Control. 2000;10(4):363-385.
75. Choi HH. On the existence of linear sliding surfaces for a class of uncertain dynamic systems with mismatched uncertainties. Automatica.
1999;35(10):1707-1715.
76. Li X, DeCarlo RA, Corless M. Sliding manifold design for a class of uncertain time delay systems - Volume 2. In: Proceedings of the 2003
American Control Conference; 2003; Denver, CO.
77. Edwards C. A practical method for the design of sliding mode controllers using linear matrix inequalities. Paper presented at: 42nd IEEE
International Conference on Decision and Control; 2003; Maui, HI.
78. Pai MC. Discrete-time sliding mode control for uncertain systems with state and input delays. Int J Syst Sci. 2010;41(12):1501-1510.
79. Nizar A, Saïd NA. A new sliding surface for discrete second order sliding mode control for time delay systems with time varying
uncertainties. Int J Comput Appl. 2012;52(21):16-22.
80. Sharav-Schapiro N, Palmor ZJ, Steinberg A. Output stabilizing robust control for discrete uncertain systems. Automatica.
1998;34(6):731-739.
81. Johansson KH, Nunes JLR. A multivariable laboratory process with an adjustable zero. In: Proceedings of the 1998 American Control
Conference - Volume 4 (ACC); 1998; Philadelphia, PA.
82. Sira-RamÍrez H. Dynamical sliding mode control strategies in the regulation of nonlinear chemical processes. Int J Control.
1992;56(1):1-21.
83. Shkolnikov IA, Shtessel YB. Nonminimum phase tracking in MIMO systems with square input-output dynamics via dynamic sliding
manifolds. J Franklin Inst. 2000;337(1):43-56.
84. Isidori A. Nonlinear Control Systems: An Introduction. 2nd ed. Berlin, Germany: Springer-Verlag Berlin Heidelberg; 1989.
85. Lee HG, Arapostathis A, Marcus SI. Linearization of discrete-time systems. Int J Control. 1987;45(5):1803-1822.
86. Xi Z, Hesketh T. On discrete time terminal sliding mode control for nonlinear systems with uncertainty. In: Proceedings of the 2010
American Control Conference; 2010; Baltimore, MD.
87. Daoutidis P, Kravaris C. Inversion and zero dynamics in nonlinear multivariable control. AICHE J. 1991;37(4):527-538.
88. Chen G, Chen N. Delay dependent decentralized H∞ output feedback control for interconnected large-scale systems with uncertainty.
Paper presented at: 2009 7th Asian Control Conference; 2009; Hong Kong, China.
89. Chen CT, Peng ST. A nonlinear control scheme for imprecisely known processes using the sliding mode and neural fuzzy techniques.
J Process Control. 2004;14(5):501-515.
90. Spurgeon S, Pugh AC. On output deadbeat control of discrete-time multivariable systems. IEEE Trans Autom Control. 1991;36(7):894-896.
91. Aliakbari S, Ayati M, Osman JH, Md Sam Y. Second-order sliding mode fault-tolerant control of heat recovery steam generator boiler in
combined cycle power plants. Appl Therm Eng. 2013;50(1):1326-1338.
40 LUIS-DELGADO ET AL.

92. Al-Hadithi BM, Jiménez A, Matía F. Variable structure control with chattering reduction of a generalized T-S model. Asian J Control.
2013;15(1):155-168.
93. Al-Hadithi BM, Barragán AJ, Andújar JM, Jiménez A. Variable structure control with chattering elimination and guaranteed stability
for a generalized T-S model. Appl Soft Comput. 2013;13(12):4802-4812.
94. Xie LB, Shieh LS, Wu CY, Tsai JSH, Canelon JI, Singla M. Digital sliding mode controller design for multiple time-delay continuous-time
transfer function matrices with a long input-output delay. J Process Control. 2015;25:78-93.
95. Xue D, Chen Y. Solving Applied Mathematical Problems With MATLAB. Boca Raton, FL: CRC Press; 2008.
96. Chen CT, Peng ST. A sliding mode control scheme for non-minimum phase non-linear uncertain input-delay chemical processes.
J Process Control. 2006;16(1):37-51.
97. Saji KS, Kumar MS. Fuzzy sliding mode control for a pH process. Paper presented at: 2010 International Conference on Communication
Control and Computing Technologies (ICCCCT); 2010; Nagercoil, India.
98. Mudi RK, Dey C, Lee TT. An improved auto-tuning scheme for PI controllers. ISA Trans. 2008;47(1):45-52.
99. Iqbal M, Bhatti AI, Iqbal S, Khan Q, Kazmi IH. Parameter estimation of uncertain nonlinear MIMO three tank systems using higher
order sliding modes. Paper presented at: 2009 IEEE International Conference on Control and Automation (ICCA); 2009; Christchurch,
New Zealand.
100. Khan Q, Bhatti AI, Iqbal S, Iqbal M. Dynamic integral sliding mode for MIMO uncertain nonlinear systems. Int J Control Autom Syst.
2011;9(1):151-160.
101. Join C, Sira-Ramírez H, Fliess M. Control of an uncertain three-tank-system via on-line parameter identification and fault detection.
IFAC Proc Vol; 2005;38(1):251-256.
102. Kovács L, Borbély E, Benyó Z. Optimal control of the three tank system in H2 /H∞ space. Paper presented at: 5th Slovakian-Hungarian
Joint Symposium on Applied Machine Intelligence and Informatics; 2007; Poprad, Slovakia.
103. Janardhanan S, Bandyopadhyay B. On discretization of continuous-time terminal sliding mode. IEEE Trans Autom Control.
2006;51(9):1532-1536.
104. Edwards C, Lai NO, Spurgeon SK. On discrete dynamic output feedback min-max controllers. Automatica. 2005;41(10):1783-1790.

How to cite this article: Luis-Delgado JD, Al-Hadithi BM, Jiménez A. Discrete-time systems sliding mode
switching hyperplane design: A survey. Optim Control Appl Meth. 2019;1–40. https://doi.org/10.1002/oca.2496

You might also like